Skip to main content

Agent Kernel

PyPI version Python 3.12+

Agent Kernel is a lightweight AI agent runtime and adapter layer for building and running AI agents across multiple frameworks. Migrate your existing agents to Agent Kernel and instantly utilize pre-built execution and testing capabilities. Deploy the same agent code without modification — see the "Multi-Cloud Deployment" section below for supported platforms.

Features

  • Unified API: Common abstractions (Agent, Runner, Session, Module, Runtime) across frameworks
  • Multi-Framework Support: OpenAI Agents SDK, CrewAI, LangGraph, Google ADK, Smolagents, Pydantic AI, Microsoft Agent Framework, and the Claude Agent SDK
  • Session Management: Built-in session abstraction with pluggable storage backends
  • Knowledge Bases: Unified, capability-declaring KnowledgeBase interface with ChromaDB, Neo4j, Starburst/Trino and Open Knowledge Format bundle (local directory or S3) backends via KnowledgeBuilder
  • Sandbox: Execute agent-generated code and shell commands in an isolated, permission-bounded environment with pluggable providers (local_subprocess, docker, kubernetes, e2b, daytona, ec2_ssm), workload profiles, policy enforcement, per-user identity, and a queue-decoupled broker for long-running executions
  • Identity and Access: One configured identity provider (any OIDC issuer such as Auth0, Keycloak or Okta, Amazon Cognito, Microsoft Entra ID, Firebase, a static credential map, or your own by dotted path) authenticates every surface; tools, hooks, threads and scheduled tasks see the same Principal, and a role mapper derives roles from groups and claims
  • Secret Resolution: SecretManager resolves API keys and passwords environment-first, falling back to a pluggable provider (env, aws_ssm for AWS SSM Parameter Store, kubernetes for Kubernetes Secrets mounted into the pod, or your own) with a TTL'd process cache
  • Scheduled Tasks: Deferred and recurring chat execution (schedule.at/schedule.cron) with a management REST API, five agent-facing tools, and pluggable provider (local, eventbridge) and store (in_memory, redis, valkey, dynamodb) backends
  • Human in the Loop: Pause a run for a person's approval or answer and resume from their decision — status: "PAUSED" with HTTP 202, a resume block carrying approved/denied/cancelled decisions, supported on OpenAI Agents SDK, LangGraph, Pydantic AI, Google ADK and Microsoft Agent Framework over REST, streaming and AG-UI
  • Flexible Deployment: Interactive CLI, REST API, serverless, or containerized deployment — see the "Multi-Cloud Deployment" section below
  • Pluggable Architecture: Easy to extend with custom framework adapters
  • MCP Server: Built-in Model Context Protocol server for exposing agents as MCP tools and exposing any custom tool
  • A2A Server: Built-in Agent-to-Agent communication server for exposing agents with a simple configuration change
  • AG-UI Server: Built-in AG-UI protocol handler for driving any streaming-capable agent from an AG-UI frontend, with opt-in shared state and client-context tools
  • A2UI: Label a JSON reply as A2UI from two config keys, so a client receives a generative-UI document as an object with its format named rather than a string to parse twice
  • REST API: Built-in REST API server for agent interaction
  • Test Automation: Built-in test suite for testing agents

Installation

uv init --python 3.12 my-agent
cd my-agent
uv add agentkernel

Install optional knowledge base extras as needed:

uv add "agentkernel[chromadb]"
uv add "agentkernel[neo4j]"
uv add "agentkernel[trino]"

Open Knowledge Format bundles need no extra - pyyaml is a core dependency. Serving a bundle from S3 uses the aws extra: uv add "agentkernel[aws]".

For LLM-based thread naming with Conversation Thread Support:

uv add "agentkernel[thread]"

For the AG-UI server (AGUIRequestHandler):

uv add "agentkernel[agui]"

For the sandbox providers (the local_subprocess provider needs no extra; ec2_ssm rides the aws extra):

uv add "agentkernel[sandbox-docker]"   # docker provider
uv add "agentkernel[kubernetes]"       # kubernetes provider (pod per sandbox)
uv add "agentkernel[e2b]"              # e2b cloud provider
uv add "agentkernel[daytona]"          # daytona cloud provider
uv add "agentkernel[aws]"              # ec2_ssm provider (boto3)

For the aws_ssm secret provider (AWS SSM Parameter Store):

uv add "agentkernel[aws]"

For cron parsing with the Scheduling capability (the eventbridge provider rides the aws extra):

uv add "agentkernel[cron]"

Requirements:

  • Python 3.12+

Quick Start

Basic Concepts

  • Agent: Framework-specific agent wrapped by an Agent Kernel adapter
  • Runner: Framework-specific execution strategy
  • Session: Shared state across conversation turns
  • Module: Container that registers agents with the Runtime
  • Runtime: Global registry and orchestrator for agents

CrewAI Example

from crewai import Agent as CrewAgent
from agentkernel.cli import CLI
from agentkernel.crewai import CrewAIModule

general_agent = CrewAgent(
    role="general",
    goal="Agent for general questions",
    backstory="You provide assistance with general queries. Give direct and short answers",
    verbose=False,
)

math_agent = CrewAgent(
    role="math",
    goal="Specialist agent for math questions",
    backstory="You provide help with math problems. Explain your reasoning at each step and include examples. \
        If prompted for anything else you refuse to answer.",
    verbose=False,
)

# Register agents with Agent Kernel
CrewAIModule([general_agent, math_agent])

if __name__ == "__main__":
    CLI.main()

LangGraph Example

from langgraph.graph import StateGraph
from agentkernel.cli import CLI
from agentkernel.langgraph import LangGraphModule

# Build and compile your graph
sg = StateGraph(...)
compiled = sg.compile()
compiled.name = "assistant"

LangGraphModule([compiled])

if __name__ == "__main__":
    CLI.main()

OpenAI Agents SDK Example

from agents import Agent as OpenAIAgent
from agentkernel.cli import CLI
from agentkernel.openai import OpenAIModule

general_agent = OpenAIAgent(
    name="general",
    handoff_description="Agent for general questions",
    instructions="You provide assistance with general queries. Give short and direct answers.",
)

OpenAIModule([general_agent])

if __name__ == "__main__":
    CLI.main()

Google ADK Example

from google.adk.agents import Agent
from agentkernel.cli import CLI
from agentkernel.adk import GoogleADKModule
from google.adk.models.lite_llm import LiteLlm

# Create Google ADK agents
math_agent = Agent(
    name="math",
    model=LiteLlm(model="openai/gpt-4o-mini"),
    description="Specialist agent for math questions",
    instruction="""
    You provide help with math problems.
    Explain your reasoning at each step and include examples.
    If prompted for anything else you refuse to answer.
    """,
)

GoogleADKModule([math_agent])

if __name__ == "__main__":
    CLI.main()

Interactive CLI

Agent Kernel includes an interactive CLI for local development and testing.

Available Commands:

  • !h, !help — Show help
  • !ld, !load <module_name> — Load a Python module containing agents
  • !ls, !list — List registered agents
  • !s, !select <agent_name> — Select an agent
  • !n, !new — Start a new session
  • !q, !quit — Exit

Usage:

uv run python demo.py

Then interact with your agents:

(assistant) >> !load my_agents
(assistant) >> !select researcher
(researcher) >> What is the latest news on AI?

Multi-Cloud Deployment

Supported Cloud Platforms: AWS, Azure, GCP

Deploy your agents to AWS, Azure, or GCP using the built-in cloud deployment handlers.

AWS Lambda Deployment

Deploy your agents as serverless functions using the built-in Lambda handler.

from openai import OpenAI
from agents import Agent as OpenAIAgent
from agentkernel.aws import Lambda
from agentkernel.openai import OpenAIModule

client = OpenAI()
assistant = OpenAIAgent(name="assistant")

OpenAIModule([assistant])
handler = Lambda.handler

Note that this is just the simple serverless version. A more advanced serverless deployment mode, which uses queues for scalability, is also available. For queue-backed execution modes and response-store configuration, see the AWS Serverless Deployment guide.

The AWS serverless handler accepts both a direct BaseRunRequest payload and the normalized BaseRequest envelope. If a flat run payload is provided, Agent Kernel generates a request_id and normalizes it before processing. The caller is never named in the payload: the API Gateway authorizer (handler = APIGatewayAuthorizer().handle, configured by the identity block) establishes the principal and the handler puts it on the request; see Identity and Access.

Accepted payloads:

{
  "prompt": "Hello agent",
  "agent": "assistant",
  "session_id": "user-123"
}
{
  "request_id": "req-123",
  "body": {
    "prompt": "Hello agent",
    "agent": "assistant",
    "session_id": "user-123"
  }
}

Azure Functions Deployment

Deploy your agents as Azure Functions using the built-in Azure handler.

from openai import OpenAI
from agents import Agent as OpenAIAgent
from agentkernel.azure import AzureFunctions
from agentkernel.openai import OpenAIModule

client = OpenAI()
assistant = OpenAIAgent(name="assistant")

OpenAIModule([assistant])
handler = AzureFunctions.handler

Request Format:

{
  "request_id": "req-123",
  "body": {
    "prompt": "Hello agent",
    "agent": "assistant",
    "session_id": "user-123"
  }
}

Azure Functions also accepts the normalized envelope, and flat run payloads are normalized in the same way before the request reaches the agent runtime.

Response Format:

{
  "result": "Agent response here"
}

Status Codes:

  • 200 — Success
  • 400 — No agent available
  • 500 — Unexpected error

GCP Cloud Run Deployment

Deploy your agents to GCP Cloud Run using the built-in CloudRun handler.

from agentkernel.gcp import CloudRun
from agentkernel.openai import OpenAIModule

OpenAIModule([...])

@CloudRun.register("/app", method="GET")
def app_handler() -> dict:
    return {"status": "ok"}

def main() -> None:
    CloudRun.run()

if __name__ == "__main__":
    main()

CloudRun is the GCP equivalent of Lambda (AWS) and AzureFunctions (Azure). It wraps RESTAPI and starts a FastAPI/uvicorn server. Custom routes are registered with @CloudRun.register(path, method). Use CloudRun.run() instead of RESTAPI.run() when deploying to GCP.

For full Terraform deployment configuration, see ak-deployment/ak-gcp/ or the GCP deployment docs.

On-Prem / Kubernetes Deployment

Deploy the queue pipeline to any Kubernetes cluster with the official Helm chart: an io-handler Deployment (REST API + Response Handler), an agent-runner Deployment, and an optional WebSocket gateway, over Kafka or NATS JetStream (or SQS on EKS). Your image supplies one entry file per component:

# app_io_handler.py
from agentkernel.pipeline import IOHandler

IOHandler.run()

# app_agent_runner.py
from agentkernel.openai import OpenAIModule
from agentkernel.pipeline import AgentRunner

OpenAIModule([...])
AgentRunner.run()

The chart injects broker and store connections as AK_* environment variables. See ak-deployment/ak-k8s/, the On-Prem / Kubernetes docs, and the end-to-end example at examples/k8s/openai-queue-mode.

Configuration

Agent Kernel can be configured via environment variables, .env files, or YAML/JSON configuration files.

Configuration Precedence

Values are loaded in the following order (highest precedence first):

  1. Environment variables (including variables from .env file)
  2. Configuration file (YAML or JSON)
  3. Built-in defaults

Configuration File

By default, Agent Kernel looks for ./config.yaml in the current working directory.

Override the config file path:

export AK_CONFIG_PATH_OVERRIDE=config.json
# or
export AK_CONFIG_PATH_OVERRIDE=conf/agent-kernel.yaml

Supported formats: .yaml, .yml, .json

Configuration Options

Logging Configuration

  • Field: logging.ak.level

  • Type: string

  • Default: WARNING

  • Description: Agent Kernel logger level (INFO, DEBUG, ERROR, WARNING, CRITICAL)

  • Environment Variable: AK_LOGGING__AK__LEVEL

  • Field: logging.system.level

  • Type: string

  • Default: WARNING

  • Description: System/root logger level (INFO, DEBUG, ERROR, WARNING, CRITICAL)

  • Environment Variable: AK_LOGGING__SYSTEM__LEVEL

Identity

The identity block selects the one identity provider every surface (REST, WebSocket, AG-UI, MCP, A2A, the CLI) authenticates through, and the role mapper that derives roles from what it asserts. Without the block every request runs as the anonymous principal. The thread, schedule and AG-UI handlers and the WebSocket modes refuse to start without it. Full reference: Identity and Access.

Auth0 is an oidc configuration: the tenant's issuer with the trailing slash, the API identifier as the audience, and the namespaced claim an Auth0 Action sets as the groups path:

identity:
  provider:
    type: oidc
    issuer: https://<tenant>.<region>.auth0.com/   # or the custom domain; trailing slash included, as Auth0 issues it
    audience: <the API identifier of your Auth0 API>
    claims:
      groups: https://<your namespace>/roles       # the namespaced claim an Auth0 Action sets; omit if you set none
  • Field: identity.provider.type

  • Type: string

  • Default: oidc

  • Description: oidc (any OpenID Connect issuer: Auth0, Keycloak, Okta, ...), cognito, entra, firebase, static, or a dotted path to an IdentityProvider subclass. oidc, cognito, entra and firebase need the auth extra (agentkernel[auth]).

  • Environment Variable: AK_IDENTITY__PROVIDER__TYPE

  • Field: identity.provider.issuer, identity.provider.audience

  • Type: string

  • Description: The issuer URL and the audience a token must carry. Both are required by oidc (Auth0: https://<tenant>.<region>.auth0.com/ with the trailing slash, and the API identifier); cognito and entra take the audience (the app client id, the application id) and derive the issuer; firebase derives both.

  • Environment Variable: AK_IDENTITY__PROVIDER__ISSUER, AK_IDENTITY__PROVIDER__AUDIENCE

  • Field: identity.provider.jwks_uri, identity.provider.algorithms, identity.provider.leeway, identity.provider.required_claims

  • Description: The JWK set URL (discovered from the issuer when omitted), the allowed signing algorithms (default [RS256, ES256]; an empty list, none, an unknown name or a symmetric one fails startup), the clock skew in seconds (default 30) and extra claims a token must carry.

  • Field: identity.provider.claims.subject, identity.provider.claims.groups

  • Type: string (a claim name, a namespaced name such as https://acme.com/roles, or dotted segments such as realm_access.roles)

  • Default: sub, groups

  • Description: Where the provider reads the subject and the group list.

  • Field: identity.provider.cognito (region, user_pool_id, token_use), identity.provider.entra (tenant_id), identity.provider.firebase (project_id)

  • Description: The settings of the matching built-in provider. token_use is access (default) or id; any other value fails startup.

  • Field: identity.provider.static.principals

  • Type: mapping of secret key to principal (subject, kind, groups, roles, claims)

  • Description: The static provider's credentials. Each key names a secret resolved through the secret capability when the provider is built (the environment first), so no credential value is written in the file; a key with no secret, or an empty map, fails startup. An entry's roles are issued directly; a role mapper replaces them.

  • Field: identity.role_mapper.type, identity.role_mapper.mapping, identity.role_mapper.claims, identity.role_mapper.pass_through

  • Description: Optional. mapping (default) maps group names to roles, and the string or list values of the claim paths listed in claims (default []: groups only, since a user sets many of their own claims), then appends the pass_through claim's value as roles; a dotted path selects a RoleMapper subclass, which reads settings of its own from model_extra. Without a mapper only a static entry's own roles exist. Nothing enforces roles yet.

Session Store

Configure where agent sessions are stored (supports multi-cloud storage backends).

  • Field: session.type
  • Type: string
  • Options: in_memory, redis, valkey (AWS), dynamodb (AWS), cosmosdb (Azure), firestore (GCP)
  • Default: in_memory
  • Environment Variable: AK_SESSION__TYPE
Redis Configuration

Required when session.type=redis:

  • URL

    • Field: session.redis.url
    • Default: redis://localhost:6379
    • Description: Redis connection URL. Use rediss:// for SSL
    • Environment Variable: AK_SESSION__REDIS__URL
  • TTL (Time to Live)

    • Field: session.redis.ttl
    • Default: 604800 (7 days)
    • Description: Session TTL in seconds
    • Environment Variable: AK_SESSION__REDIS__TTL
  • Key Prefix

    • Field: session.redis.prefix
    • Default: ak:sessions:
    • Description: Key prefix for session storage
    • Environment Variable: AK_SESSION__REDIS__PREFIX
Valkey Configuration

Required when session.type=valkey (requires the agentkernel[valkey] extra). Valkey is the open-source, Linux Foundation-governed fork of Redis — wire-compatible with Redis and available on AWS ElastiCache at a lower price point than the Redis OSS engine:

  • URL

    • Field: session.valkey.url
    • Default: valkey://localhost:6379
    • Description: Valkey connection URL. Use valkeys:// for SSL
    • Environment Variable: AK_SESSION__VALKEY__URL
  • TTL (Time to Live)

    • Field: session.valkey.ttl
    • Default: 604800 (7 days)
    • Description: Session TTL in seconds
    • Environment Variable: AK_SESSION__VALKEY__TTL
  • Key Prefix

    • Field: session.valkey.prefix
    • Default: ak:sessions:
    • Description: Key prefix for session storage
    • Environment Variable: AK_SESSION__VALKEY__PREFIX

Conversation Thread Support

Mounting AgentThreadRequestHandler (from agentkernel.thread) instead of the default REST handler enables persistent, named conversation threads keyed by session_id: it serves the standard chat routes with thread recording, plus the read routes (GET /api/v1/threads and GET /api/v1/threads/{session_id}). The handler requires an identity block: every thread belongs to the authenticated principal, an anonymous request is a 401, and the listing returns the caller's own threads (there is no user_id field or query parameter). The thread block in the configuration only selects the store backend and naming. A thread is auto-created on a session's first request. Sending thread_name on any chat request sets or renames the thread's display name and locks it against automatic naming. Threads created without an explicit thread_name are named by a pluggable naming strategy — by default an LLM call derives a concise title from the first prompt (falling back to a prefix of the prompt when litellm or an API key is unavailable). Attachments in thread mode additionally require multimodal.enabled: true with a shared attachment store (in_memory, redis, or dynamodb — session_cache is rejected). See examples/api/thread-openai and examples/api/multimodal/thread-openai.

  • Field: thread.type

  • Type: string

  • Options: in_memory, redis, valkey, dynamodb (AWS), firestore (GCP), cosmosdb (Azure)

  • Default: in_memory

  • Environment Variable: AK_THREAD__TYPE

  • Naming Model

    • Field: thread.naming.model
    • Type: string
    • Default: gpt-4o-mini
    • Description: LiteLLM model used to generate thread names (requires the thread extra — uv add "agentkernel[thread]" — and an API key in the environment; falls back to a truncated prompt prefix otherwise)
    • Environment Variable: AK_THREAD__NAMING__MODEL
  • Auto-name Max Length

    • Field: thread.naming.max_length
    • Type: integer
    • Default: 80
    • Description: Maximum length of an auto-generated thread name
    • Environment Variable: AK_THREAD__NAMING__MAX_LENGTH
Redis Thread Store

Required when thread.type=redis:

  • URL

    • Field: thread.redis.url
    • Default: redis://localhost:6379
    • Description: Redis connection URL. Use rediss:// for SSL
    • Environment Variable: AK_THREAD__REDIS__URL
  • TTL (Time to Live)

    • Field: thread.redis.ttl
    • Default: 2592000 (30 days)
    • Description: Thread TTL in seconds (0 disables)
    • Environment Variable: AK_THREAD__REDIS__TTL
  • Key Prefix

    • Field: thread.redis.prefix
    • Default: ak:thread:
    • Description: Key prefix for Redis thread storage
    • Environment Variable: AK_THREAD__REDIS__PREFIX
Valkey Thread Store

Required when thread.type=valkey. Requires the valkey extra (uv add "agentkernel[valkey]"):

  • URL

    • Field: thread.valkey.url
    • Default: valkey://localhost:6379
    • Description: Valkey connection URL. Use valkeys:// for SSL
    • Environment Variable: AK_THREAD__VALKEY__URL
  • TTL (Time to Live)

    • Field: thread.valkey.ttl
    • Default: 2592000 (30 days)
    • Description: Thread TTL in seconds (0 disables)
    • Environment Variable: AK_THREAD__VALKEY__TTL
  • Key Prefix

    • Field: thread.valkey.prefix
    • Default: ak:thread:
    • Description: Key prefix for Valkey thread storage
    • Environment Variable: AK_THREAD__VALKEY__PREFIX
DynamoDB Thread Store

Used when thread.type=dynamodb:

  • Table Name

    • Field: thread.dynamodb.table_name
    • Default: ak-agent-threads
    • Description: DynamoDB table name. The table must have a partition key named session_id (S) and a sort key named sk (S)
    • Environment Variable: AK_THREAD__DYNAMODB__TABLE_NAME
  • TTL (Time to Live)

    • Field: thread.dynamodb.ttl
    • Default: 0 (disabled)
    • Description: DynamoDB item TTL in seconds
    • Environment Variable: AK_THREAD__DYNAMODB__TTL
Firestore Thread Store

Used when thread.type=firestore:

  • Collection Name

    • Field: thread.firestore.collection_name
    • Default: ak-agent-threads
    • Description: Firestore collection name; each document ID is a session_id
    • Environment Variable: AK_THREAD__FIRESTORE__COLLECTION_NAME
  • Project ID

    • Field: thread.firestore.project_id
    • Default: null (inferred from Application Default Credentials)
    • Environment Variable: AK_THREAD__FIRESTORE__PROJECT_ID
  • Database ID

    • Field: thread.firestore.database_id
    • Default: null (the (default) database)
    • Environment Variable: AK_THREAD__FIRESTORE__DATABASE_ID
  • TTL (Time to Live)

    • Field: thread.firestore.ttl
    • Default: 0 (disabled)
    • Description: Thread TTL in seconds
    • Environment Variable: AK_THREAD__FIRESTORE__TTL
Cosmos DB Thread Store

Required when thread.type=cosmosdb:

  • Connection String

    • Field: thread.cosmosdb.connection_string
    • Description: Cosmos DB connection string (Azure Portal → Keys). Uses the Table API; entities are partitioned by session_id. No TTL support
    • Environment Variable: AK_THREAD__COSMOSDB__CONNECTION_STRING
  • Table Name

    • Field: thread.cosmosdb.table_name
    • Default: akagentthreads
    • Description: Cosmos DB table name for thread storage
    • Environment Variable: AK_THREAD__COSMOSDB__TABLE_NAME

Scheduling

The presence of a schedule block enables deferred and recurring chat execution: a chat request carrying a schedule block (at for one-time, cron for recurring, plus timezone and session_mode) is not run — it is registered as a scheduled task and acknowledged with HTTP 202. When an occurrence is due, the provider delivers the stored prompt into the input queue as a plain chat request, so scheduling requires the queue execution pipeline. The block also injects five agent tools (create_schedule, list_schedules, get_schedule, update_schedule, delete_schedule). The management routes (GET/PUT/DELETE /api/v1/schedules) are not mounted from config: the application mounts ScheduleRESTRequestHandler itself — IOHandler.run(handlers=[ScheduleRESTRequestHandler()]) — and the handler requires an identity block. Every scheduling request needs an authenticated principal: it is the owner the task is stored under and that later reads and changes are checked against (an anonymous management request is a 401, an anonymous chat request with a schedule block a 403; there is no user_id field). A bare schedule: block works for local development: its defaults are the local provider and the in_memory store. See examples/api/schedule-openai.

  • Provider Type

    • Field: schedule.provider.type
    • Type: string
    • Default: local
    • Options: local (in-process scheduler thread; requires the in_memory transport and store), eventbridge (AWS EventBridge Scheduler; requires the sqs transport and the aws extra), or a dotted path to a ScheduleProvider subclass
    • Environment Variable: AK_SCHEDULE__PROVIDER__TYPE
  • Store Type

    • Field: schedule.store.type
    • Type: string
    • Default: in_memory
    • Options: in_memory, redis, valkey, dynamodb, or a dotted path to a ScheduleStore subclass
    • Environment Variable: AK_SCHEDULE__STORE__TYPE
  • Tool Scoping

    • Field: schedule.agents
    • Type: list of strings
    • Default: null (all agents)
    • Description: Agent names the schedule tools and system-prompt guidance attach to
    • Environment Variable: AK_SCHEDULE__AGENTS
EventBridge Scheduler Provider

Required when schedule.provider.type=eventbridge. All three are supplied by the AWS Terraform modules when enable_scheduling = true; a missing one fails at startup with an AKConfigError.

  • Group Name

    • Field: schedule.provider.eventbridge.group_name
    • Description: EventBridge Scheduler schedule-group name the schedules are created in
    • Environment Variable: AK_SCHEDULE__PROVIDER__EVENTBRIDGE__GROUP_NAME
  • Role ARN

    • Field: schedule.provider.eventbridge.role_arn
    • Description: Execution role ARN Scheduler assumes to deliver triggers to the input queue
    • Environment Variable: AK_SCHEDULE__PROVIDER__EVENTBRIDGE__ROLE_ARN
  • Queue ARN

    • Field: schedule.provider.eventbridge.queue_arn
    • Description: Input queue ARN used as the schedule target
    • Environment Variable: AK_SCHEDULE__PROVIDER__EVENTBRIDGE__QUEUE_ARN
Redis / Valkey Schedule Store

Required when schedule.store.type=redis (or valkey, with schedule.store.valkey.* / AK_SCHEDULE__STORE__VALKEY__*).

  • URL

    • Field: schedule.store.redis.url
    • Default: redis://localhost:6379
    • Description: Redis connection URL. Use rediss:// for SSL
    • Environment Variable: AK_SCHEDULE__STORE__REDIS__URL
  • Key Prefix

    • Field: schedule.store.redis.prefix
    • Default: ak:schedule:
    • Description: Key prefix for scheduled-task storage
    • Environment Variable: AK_SCHEDULE__STORE__REDIS__PREFIX
  • TTL (Time to Live)

    • Field: schedule.store.redis.ttl
    • Default: 0 (disabled)
    • Description: Scheduled task TTL in seconds. Unlike threads this defaults to 0 — a task that silently expired would stop firing with no audit trail
    • Environment Variable: AK_SCHEDULE__STORE__REDIS__TTL
DynamoDB Schedule Store

Required when schedule.store.type=dynamodb. The table needs a partition key named task_id (S) and no sort key; the AWS Terraform modules create it when create_dynamodb_schedule_table = true.

  • Table Name

    • Field: schedule.store.dynamodb.table_name
    • Default: ak-agent-schedules
    • Description: DynamoDB table name for scheduled-task storage
    • Environment Variable: AK_SCHEDULE__STORE__DYNAMODB__TABLE_NAME
  • TTL (Time to Live)

    • Field: schedule.store.dynamodb.ttl
    • Default: 0 (disabled)
    • Description: DynamoDB item TTL in seconds
    • Environment Variable: AK_SCHEDULE__STORE__DYNAMODB__TTL

Secret Resolution

SecretManager.current().get("OPENAI_API_KEY") resolves a secret by the same environment-variable-style name the SDKs already read (^[A-Z][A-Z0-9_]*$). Resolution order is fixed: a set, non-empty environment variable always wins, then the process cache, then the configured provider. A resolved value is returned to the caller and never written back to the environment, so hand it to the SDK explicitly (set_default_openai_key(...)). A miss raises SecretNotFoundError unless a default is passed; a provider failure raises SecretError even with a default. The capability is always available — there is no enabled flag; selecting a provider other than env is the only opt-in. See examples/cli/openai-secret.

  • Provider Type

    • Field: secret.provider.type
    • Type: string
    • Default: env
    • Options: env (the process environment), aws_ssm (AWS SSM Parameter Store; requires secret.prefix and the aws extra — OPENAI_API_KEY is read from /ak/<prefix>/openai_api_key), kubernetes (one file per key from Kubernetes Secrets mounted under secret.provider.kubernetes.mount_path, standard library only, no API calls), or a dotted path to a SecretProvider subclass
    • Environment Variable: AK_SECRET__PROVIDER__TYPE
  • Kubernetes Mount Path

    • Field: secret.provider.kubernetes.mount_path
    • Type: string (absolute path)
    • Default: /var/run/secrets/agentkernel
    • Description: Directory the Kubernetes Secrets are mounted under, one subdirectory per Secret (<mount_path>/<secret-name>/<KEY>). The Helm chart's secretStore mounts under the default. Read only by the kubernetes provider
    • Environment Variable: AK_SECRET__PROVIDER__KUBERNETES__MOUNT_PATH
  • Prefix

    • Field: secret.prefix
    • Type: string
    • Default: ""
    • Description: Deployment scope providers namespace secrets under; a single path segment. Required by aws_ssm, ignored by env and kubernetes. Injected by the AWS Terraform modules when ssm_enabled = true
    • Environment Variable: AK_SECRET__PREFIX
  • Cache TTL

    • Field: secret.cache_ttl
    • Type: integer
    • Default: 300
    • Description: Seconds a provider hit is served from the process cache (0 disables caching); the rotation-pickup window for values read per call
    • Environment Variable: AK_SECRET__CACHE_TTL

Execution Configuration

Configure queue-backed and serverless execution behavior.

  • Execution Mode

    • Field: execution.mode
    • Options: rest_sync, rest_async, async, realtime
    • Default: None (runs as rest_sync)
    • Description: How a request is invoked: rest_sync (REST, wait for the reply), rest_async (REST, submit then poll), async (WebSocket only), realtime (persistent realtime sockets, e.g. the LiveKit voice gateway). stream was removed and fails at config load with AKConfigError: set mode: rest_sync with message_type: delta for SSE over REST, or mode: async with message_type: delta for chunks over WebSocket
    • Environment Variable: AK_EXECUTION__MODE
  • Message Type

    • Field: execution.message_type
    • Options: complete, delta
    • Default: None (resolves to delta for realtime, complete otherwise)
    • Description: How much of the reply each delivered message carries. complete sends the whole reply in one message (JSON under rest_sync, one CHAT_RESPONSE under async); delta sends one message per chunk (SSE under rest_sync, one STREAM_CHUNK per chunk ending with done: true under async). rest_async + delta and realtime + complete fail at config load. AWS Lambda (REST router init) and the ECS queue-mode ECSIOHandler (startup) refuse rest_sync + delta; use async + delta there
    • Environment Variable: AK_EXECUTION__MESSAGE_TYPE
  • Queues

    • Field: execution.queues

    • Description: Queue settings used by the queue execution pipeline (in-process by default, or serverless/containerized backends)

    • Transport Type

      • Field: execution.queues.type
      • Options: in_memory, sqs, kafka, nats, or a dotted path to a QueueTransport subclass
      • Default: none — mandatory whenever an execution.queues block is declared
      • Description: Queue transport used by the pipeline, and the only thing that selects it: queue coordinates are injected per component by a deployment, so they are never used to infer the transport. Declaring the block without a type is a configuration error. Omitting the block entirely leaves in_memory — a zero-dependency, in-process transport for local development and single-process deployments.
      • Environment Variable: AK_EXECUTION__QUEUES__TYPE
    • Input Queue URL

      • Field: execution.queues.input.url
      • Default: None
      • Description: Input queue URL (sqs transport only)
      • Environment Variable: AK_EXECUTION__QUEUES__INPUT__URL
    • Output Queue URL

      • Field: execution.queues.output.url
      • Default: None
      • Description: Output queue URL (sqs transport only)
      • Environment Variable: AK_EXECUTION__QUEUES__OUTPUT__URL
    • Input Queue Max Receive Count

      • Field: execution.queues.input.max_receive_count
      • Default: 3
      • Environment Variable: AK_EXECUTION__QUEUES__INPUT__MAX_RECEIVE_COUNT
    • Output Queue Max Receive Count

      • Field: execution.queues.output.max_receive_count
      • Default: 3
      • Environment Variable: AK_EXECUTION__QUEUES__OUTPUT__MAX_RECEIVE_COUNT
    • Input Queue Consumer Count

      • Field: execution.queues.input.no_of_consumers
      • Default: 5
      • Description: Number of independent consumer threads that each poll the input queue in a continuous loop. Used by the in-process pipeline (agent-runner worker threads) and by ECS containerized deployments; not used in serverless (Lambda) mode, which has no consumer threads.
      • Environment Variable: AK_EXECUTION__QUEUES__INPUT__NO_OF_CONSUMERS
    • Output Queue Consumer Count

      • Field: execution.queues.output.no_of_consumers
      • Default: 5
      • Description: Number of independent consumer threads that each poll the output queue in a continuous loop. Used by the in-process pipeline (response-handler worker threads) and by ECS containerized deployments; not used in serverless (Lambda) mode, which has no consumer threads.
      • Environment Variable: AK_EXECUTION__QUEUES__OUTPUT__NO_OF_CONSUMERS
    • In-Memory Transport Ack Wait

      • Field: execution.queues.in_memory.ack_wait
      • Default: 300.0
      • Description: Seconds an unacknowledged in-memory message stays invisible before redelivery. Redelivery rescues stuck worker threads; keep this above your longest expected agent run or a slow run will be executed again.
    • In-Memory Transport Dedup Window

      • Field: execution.queues.in_memory.dedup_window
      • Default: 300.0
      • Description: Seconds within which a repeated message_deduplication_id is dropped
    • Kafka Bootstrap Servers

      • Field: execution.queues.kafka.bootstrap_servers
      • Default: localhost:9092
      • Description: Kafka bootstrap servers (host:port, comma-separated)
      • Environment Variable: AK_EXECUTION__QUEUES__KAFKA__BOOTSTRAP_SERVERS
    • Kafka Input Topic

      • Field: execution.queues.kafka.input_topic
      • Default: agent-input
      • Description: Topic carrying chat requests
      • Environment Variable: AK_EXECUTION__QUEUES__KAFKA__INPUT_TOPIC
    • Kafka Output Topic

      • Field: execution.queues.kafka.output_topic
      • Default: agent-output
      • Description: Topic carrying agent replies
      • Environment Variable: AK_EXECUTION__QUEUES__KAFKA__OUTPUT_TOPIC
    • Kafka Consumer Group Id

      • Field: execution.queues.kafka.group_id
      • Default: agent-kernel
      • Description: Consumer group id prefix; the input and output consumers append their queue name to it
      • Environment Variable: AK_EXECUTION__QUEUES__KAFKA__GROUP_ID
    • Kafka Dead-Letter Topic Suffix

      • Field: execution.queues.kafka.dlq_suffix
      • Default: .dlq
      • Description: Suffix appended to a topic name for its dead-letter topic, where permanently failed records are routed
      • Environment Variable: AK_EXECUTION__QUEUES__KAFKA__DLQ_SUFFIX
    • Kafka Retry Backoff

      • Field: execution.queues.kafka.retry_backoff
      • Default: 2.0
      • Description: Seconds to wait before an in-process retry of a failed record
      • Environment Variable: AK_EXECUTION__QUEUES__KAFKA__RETRY_BACKOFF
    • Kafka Delivery Timeout

      • Field: execution.queues.kafka.delivery_timeout
      • Default: 30.0
      • Description: Seconds to wait for the broker to confirm a produced message before failing the send
      • Environment Variable: AK_EXECUTION__QUEUES__KAFKA__DELIVERY_TIMEOUT
    • Kafka Metadata Timeout

      • Field: execution.queues.kafka.metadata_timeout
      • Default: 5.0
      • Description: Seconds to wait for topic metadata during the startup partition-capacity check (the check is skipped on timeout)
      • Environment Variable: AK_EXECUTION__QUEUES__KAFKA__METADATA_TIMEOUT
    • Kafka Client Config

      • Field: execution.queues.kafka.client_config
      • Default: {}
      • Description: Passthrough settings merged into the confluent-kafka producer and consumer configs (SASL, TLS, tuning); set via config.yaml, not individually exported as environment variables
    • NATS Server URL

      • Field: execution.queues.nats.url
      • Default: nats://localhost:4222
      • Description: NATS server URL (comma-separated for a cluster)
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__URL
    • NATS Input Stream

      • Field: execution.queues.nats.input_stream
      • Default: AGENT_REQUESTS
      • Description: JetStream stream carrying chat requests
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__INPUT_STREAM
    • NATS Input Subject Prefix

      • Field: execution.queues.nats.input_subject_prefix
      • Default: chat.req
      • Description: Subject prefix for chat requests
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__INPUT_SUBJECT_PREFIX
    • NATS Output Stream

      • Field: execution.queues.nats.output_stream
      • Default: AGENT_REPLIES
      • Description: JetStream stream carrying agent replies
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__OUTPUT_STREAM
    • NATS Output Subject Prefix

      • Field: execution.queues.nats.output_subject_prefix
      • Default: chat.out
      • Description: Subject prefix for agent replies
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__OUTPUT_SUBJECT_PREFIX
    • NATS Partitions

      • Field: execution.queues.nats.partitions
      • Default: 32
      • Description: Number of partition subjects per stream, each served by its own durable consumer. Sessions hash to a partition, so this caps how many messages can be in flight at once: keep it at or above no_of_consumers x replicas. Changing it re-maps sessions, so size it up front (idle partitions cost almost nothing)
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__PARTITIONS
    • NATS Ack Wait

      • Field: execution.queues.nats.ack_wait
      • Default: 300.0
      • Description: Seconds the server waits for an acknowledgement before redelivering. Must exceed your longest agent turn: a turn that outlives it is redelivered and executed a second time
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__ACK_WAIT
    • NATS Retry Backoff

      • Field: execution.queues.nats.retry_backoff
      • Default: 2.0
      • Description: Seconds to delay a redelivery after a failed message (nak delay)
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__RETRY_BACKOFF
    • NATS Duplicate Window

      • Field: execution.queues.nats.duplicate_window
      • Default: 300.0
      • Description: Seconds within which a repeated dedup id is dropped by the stream (SQS parity)
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__DUPLICATE_WINDOW
    • NATS Max Age

      • Field: execution.queues.nats.max_age
      • Default: 86400.0
      • Description: Seconds before an unconsumed message is discarded. A safety net: work-queue messages are otherwise kept forever
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__MAX_AGE
    • NATS Request Timeout

      • Field: execution.queues.nats.request_timeout
      • Default: 10.0
      • Description: Seconds to wait for a NATS request (publish, ack, management call) to complete
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__REQUEST_TIMEOUT
    • NATS Auto Provision

      • Field: execution.queues.nats.auto_provision
      • Default: false
      • Description: Create the streams and per-partition consumers at startup if missing. Convenient for local and dev clusters; leave false in production, where the objects are managed declaratively (NACK CRs) and a missing object should fail loudly instead of being created with defaults
      • Environment Variable: AK_EXECUTION__QUEUES__NATS__AUTO_PROVISION
    • Queue Batch Size

      • Field: execution.queues.batch_size
      • Default: None
      • Description: Max number of messages fetched per receive call, shared by the input and output queues. Only used by containerized deployments — never set for serverless deployments, which control batch size differently. Controlled by the deployment tooling via env var AK_EXECUTION__QUEUES__BATCH_SIZE — do not set in config.yaml.
      • Environment Variable: AK_EXECUTION__QUEUES__BATCH_SIZE
  • Response Store

    • Field: execution.response_store

    • Description: Response persistence settings used by the serverless response handler

    • Type

      • Field: execution.response_store.type
      • Options: in_memory, redis, valkey, dynamodb, or a dotted path to a ResponseStore subclass
      • Description: Response store backend selector configured in config.yaml; this value is not exported as an environment variable. Defaults to an in-process in_memory store when unset and no other backend is configured.
    • Retry Count

      • Field: execution.response_store.retry_count
      • Default: 5
      • Description: Number of lookup attempts when polling for a response
      • Environment Variable: AK_EXECUTION__RESPONSE_STORE__RETRY_COUNT
    • Delay

      • Field: execution.response_store.delay
      • Default: 5
      • Description: Delay in seconds between response lookup attempts
      • Environment Variable: AK_EXECUTION__RESPONSE_STORE__DELAY
    • Redis Backend

      • Field: execution.response_store.redis
      • Environment Variables: AK_EXECUTION__RESPONSE_STORE__REDIS__URL, AK_EXECUTION__RESPONSE_STORE__REDIS__PREFIX, AK_EXECUTION__RESPONSE_STORE__REDIS__TTL
    • Valkey Backend

      • Field: execution.response_store.valkey
      • Environment Variables: AK_EXECUTION__RESPONSE_STORE__VALKEY__URL, AK_EXECUTION__RESPONSE_STORE__VALKEY__PREFIX, AK_EXECUTION__RESPONSE_STORE__VALKEY__TTL
      • Description: Valkey-backed response storage (requires the agentkernel[valkey] extra)
    • DynamoDB Backend

      • Field: execution.response_store.dynamodb
      • Environment Variables: AK_EXECUTION__RESPONSE_STORE__DYNAMODB__TABLE_NAME, AK_EXECUTION__RESPONSE_STORE__DYNAMODB__TTL
      • Description: DynamoDB-backed response storage with table name and TTL

Use the built-in in_memory response store for local development and single-process deployments (zero extra services required), or Redis, Valkey, or DynamoDB for distributed/serverless deployments. The runtime accepts BaseRunRequest payloads directly, normalizes them internally when queueing is required, and uses request_id plus optional user_id as queue message attributes.

API Configuration

Configure the REST API server (if using the API module).

  • Host

    • Field: api.host
    • Default: 0.0.0.0
    • Environment Variable: AK_API__HOST
  • Port

    • Field: api.port
    • Default: 8000
    • Environment Variable: AK_API__PORT
  • Custom Router Prefix

    • Field: api.custom_router_prefix
    • Default: /custom
    • Environment Variable: AK_API__CUSTOM_ROUTER_PREFIX
  • Enabled Routes

    • Field: api.enabled_routes.agents
    • Default: true
    • Description: Enable agent interaction routes
    • Environment Variable: AK_API__ENABLED_ROUTES__AGENTS

A2A (Agent-to-Agent) Configuration

  • Enabled

    • Field: a2a.enabled
    • Default: false
    • Environment Variable: AK_A2A__ENABLED
  • Agents

    • Field: a2a.agents
    • Default: ["*"]
    • Description: List of agent names to enable A2A (use ["*"] for all)
    • Environment Variable: AK_A2A__AGENTS (comma-separated)
  • URL

    • Field: a2a.url
    • Default: http://localhost:8000/a2a
    • Environment Variable: AK_A2A__URL
  • Task Store Type

    • Field: a2a.task_store_type
    • Options: in_memory, redis
    • Default: in_memory
    • Environment Variable: AK_A2A__TASK_STORE_TYPE

MCP (Model Context Protocol) Configuration

  • Enabled

    • Field: mcp.enabled
    • Default: false
    • Environment Variable: AK_MCP__ENABLED
  • Expose Agents

    • Field: mcp.expose_agents
    • Default: false
    • Description: Expose agents as MCP tools
    • Environment Variable: AK_MCP__EXPOSE_AGENTS
  • Agents

    • Field: mcp.agents
    • Default: ["*"]
    • Description: List of agent names to expose as MCP tools
    • Environment Variable: AK_MCP__AGENTS (comma-separated)
  • URL

    • Field: mcp.url
    • Default: http://localhost:8000/mcp
    • Description: Public URL of the MCP endpoint. With an OpenID-family identity provider the 401 points at the RFC 9728 resource metadata the app serves at <origin>/.well-known/oauth-protected-resource/mcp/, naming this URL as the resource; set it to the address clients use (behind a proxy, the public one)
    • Environment Variable: AK_MCP__URL
  • Stateless HTTP

    • Field: mcp.stateless_http
    • Default: false
    • Description: Run MCP in stateless HTTP mode (no Mcp-Session-Id)
    • Environment Variable: AK_MCP__STATELESS_HTTP
  • Endpoint (not configurable)

    • The MCP server is always mounted at /mcp on the main API server.
    • Full URL: http://{api.host}:{api.port}/mcp — use api.port / AK_API__PORT to change the port.

AG-UI Configuration

Mounting AGUIRequestHandler (from agentkernel.agui, requires the agentkernel[agui] extra — pip install "agentkernel[agui]") is what enables the AG-UI protocol surface; the agui block only parameterizes it, and it never switches the surface on by itself. Mount AGUIPipelineRequestHandler through IOHandler.run(handlers=[...]) instead to run the agent on the queue pipeline rather than inside the SSE request; same routes and events, and it refuses to start unless the configured stores can serve a run from the runner process. AGUIRequestHandler refuses to construct without an identity block and answers an anonymous request with a 401, since AG-UI runs agents on a caller's behalf and has no anonymous mode. Only agents whose runner declares supports_streaming = True are reachable (currently OpenAI Agents SDK, LangGraph, Google ADK, Pydantic AI and Microsoft Agent Framework — not CrewAI or Smolagents). See examples/api/agui.

  • Agents

    • Field: agui.agents
    • Default: unset (every streaming-capable agent is reachable)
    • Description: Agent names reachable over AG-UI
    • Environment Variable: AK_AGUI__AGENTS (comma-separated)
  • Prefix

    • Field: agui.prefix
    • Default: /agui
    • Description: Route prefix for the AG-UI surface
    • Environment Variable: AK_AGUI__PREFIX
  • Default Agent

    • Field: agui.default_agent
    • Default: unset
    • Description: Agent served on the bare prefix route (POST {prefix}, in addition to POST {prefix}/{agent_name}); must be one of agui.agents when that list is set
    • Environment Variable: AK_AGUI__DEFAULT_AGENT
  • State Tools

    • Field: agui.state.enabled
    • Default: false
    • Description: Attach get_agui_state / update_agui_state, giving agents read/write access to AG-UI's shared JSON state (a StateSnapshot is streamed back only when the state actually changed)
    • Environment Variable: AK_AGUI__STATE__ENABLED
    • Field: agui.state.agents
    • Default: unset (every agent gets the tools)
    • Environment Variable: AK_AGUI__STATE__AGENTS (comma-separated)
  • Client Context Tools

    • Field: agui.client_context.enabled
    • Default: false
    • Description: Attach the read-only get_forwarded_props / get_agui_context tools over a run's forwardedProps and context fields; never injected into the prompt automatically
    • Environment Variable: AK_AGUI__CLIENT_CONTEXT__ENABLED
    • Field: agui.client_context.agents
    • Default: unset (every agent gets the tools)
    • Environment Variable: AK_AGUI__CLIENT_CONTEXT__AGENTS (comma-separated)

A2UI Configuration

Labels a JSON reply as A2UI so a client receives the document as an object with its format named, instead of a JSON string it has to parse a second time. Agent Kernel supplies the media type and nothing else — the component catalog belongs in your agent's instructions, because it has to match the components your frontend implements. The block does not reach a streamed run (execution.message_type: delta) or an agent configured with a framework output_type; both are told to write their own post-hook. See examples/api/a2ui-openai.

  • Enabled

    • Field: a2ui.enabled
    • Default: false
    • Description: Master switch. When off, no hook runs and no reply is labelled
    • Environment Variable: AK_A2UI__ENABLED
  • Agents

    • Field: a2ui.agents
    • Default: unset (every agent's JSON replies are labelled)
    • Description: Agent names whose JSON replies are labelled application/a2ui+json. Naming an agent is the declaration that it emits A2UI; a reply from it that does not parse as JSON is left alone
    • Environment Variable: AK_A2UI__AGENTS (comma-separated)

Trace (Observability) Configuration

Configure tracing and observability for monitoring agent execution.

  • Enabled

    • Field: trace.enabled
    • Default: false
    • Description: Enable tracing/observability
    • Environment Variable: AK_TRACE__ENABLED
  • Type

    • Field: trace.type
    • Options: langfuse, openllmetry, logfire, cloudwatch
    • Default: langfuse
    • Description: Type of tracing provider to use
    • Environment Variable: AK_TRACE__TYPE

Langfuse Setup:

To use Langfuse for tracing, install the langfuse extra:

uv add "agentkernel[langfuse]"

Configure Langfuse credentials via environment variables:

export LANGFUSE_PUBLIC_KEY=pk-lf-...
export LANGFUSE_SECRET_KEY=sk-lf-...
export LANGFUSE_HOST=https://cloud.langfuse.com  # or your self-hosted instance

Enable tracing in your configuration:

trace:
  enabled: true
  type: langfuse

OpenLLMetry (Traceloop) Setup:

To use OpenLLMetry for tracing, install the openllmetry extra:

uv add "agentkernel[openllmetry]"

Configure Traceloop credentials via environment variables:

export TRACELOOP_API_KEY=your-api-key
export TRACELOOP_BASE_URL=https://api.traceloop.com  # Optional: for self-hosted

Enable tracing in your configuration:

trace:
  enabled: true
  type: openllmetry

Pydantic Logfire Setup:

To use Logfire for tracing, install the logfire extra:

uv add "agentkernel[logfire]"

Configure the Logfire write token via an environment variable (optional — without a token, Logfire runs locally and does not ship traces):

export LOGFIRE_TOKEN=your-write-token

Enable tracing in your configuration:

trace:
  enabled: true
  type: logfire

AWS CloudWatch Setup:

To send traces to Amazon CloudWatch (via AWS X-Ray, searchable in Transaction Search), install the cloudwatch extra:

uv add "agentkernel[cloudwatch]"

Spans are exported over OTLP to the X-Ray endpoint of your region, signed with the standard AWS credential chain (environment, profile, SSO, or instance/task role). Enable CloudWatch Transaction Search once in the account, and give the agent's role the AWSXrayWriteOnlyAccess managed policy:

export AWS_REGION=us-east-1
# Optional: export through the CloudWatch agent or an OpenTelemetry collector instead
# export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=http://localhost:4318/v1/traces

Enable tracing in your configuration:

trace:
  enabled: true
  type: cloudwatch

Fetching spans back (trace evaluation):

Every built-in tracer can read recorded spans back through one provider-neutral filter, so production traces can be evaluated offline:

from datetime import timedelta

from agentkernel.trace import SpanKind, Trace

tool_calls = Trace.get().fetch_spans(last=timedelta(hours=6), kinds=[SpanKind.TOOL])
traces = Trace.get().fetch_traces(session_id="session-123", last=timedelta(hours=24))

Spans are returned oldest first as FetchedSpan objects. See the Traceability docs and the examples/aws-serverless/openai-langfuse-trace-evaluation example. Custom tracers that override fetch_spans can be checked with agentkernel.trace.testing.TraceFetchContract.

Test Configuration

Configure test comparison modes for automated testing. Test configuration is separate from the application configuration: it is not part of config.yaml. It lives in its own test-config.yaml file and is only loaded when the testing utilities (agentkernel.test) are used — see the Test Configuration (test-config.yaml) section for file resolution, environment variables, and migration notes.

  • Mode

    • Field: mode
    • Options: score, llm, fallback
    • Default: fallback
    • Description: Test comparison mode
    • Environment Variable: AK_TEST__MODE
  • Evaluator

    • Field: evaluator
    • Default: deepeval
    • Description: Built-in evaluator short name (deepeval, opik or jev), or a dotted path to your own AKEvaluator subclass
    • Environment Variable: AK_TEST__EVALUATOR
  • Llm Model

    • Field: llm.model
    • Default: gpt-4o-mini
    • Description: LLM model for llm evaluation
    • Environment Variable: AK_TEST__LLM__MODEL
  • Llm Provider

    • Field: llm.provider
    • Default: openai
    • Description: LLM provider for llm evaluation
    • Environment Variable: AK_TEST__LLM__PROVIDER
  • Llm Embedding Model

    • Field: llm.embedding_model
    • Default: text-embedding-3-small
    • Description: Embedding model, unconsumed by any built-in v1 metric
    • Environment Variable: AK_TEST__LLM__EMBEDDING_MODEL

Test Modes:

  • score: Deterministic, offline string-match scoring via the configured evaluator (built-in deepeval: Scorer.quasi_exact_match_score; built-in opik: LevenshteinRatio, requires uv add "agentkernel[opik]"; built-in jev has no score mode)
  • llm: LLM-as-judge evaluation via the configured evaluator (built-in deepeval or opik: GEval; built-in jev: a TypeSafe JEV yes/no Noul question) for semantic similarity
  • fallback: Tries score first, falls back to llm if score fails
# test-config.yaml (separate file — not config.yaml)
mode: fallback
evaluator: deepeval
llm:
  model: gpt-4o-mini
  provider: openai
  embedding_model: text-embedding-3-small

Guardrails Configuration

Configure input and output guardrails to validate agent requests and responses for safety and compliance.

  • Input Guardrails

    • Enabled

      • Field: guardrail.input.enabled
      • Default: false
      • Description: Enable input validation guardrails
      • Environment Variable: AK_GUARDRAIL__INPUT__ENABLED
    • Type

      • Field: guardrail.input.type
      • Default: openai
      • Options: openai, bedrock, walledai
      • Description: Guardrail provider type
      • Environment Variable: AK_GUARDRAIL__INPUT__TYPE
    • Config Path

      • Field: guardrail.input.config_path
      • Default: None
      • Description: Path to guardrail configuration JSON file (OpenAI only)
      • Environment Variable: AK_GUARDRAIL__INPUT__CONFIG_PATH
    • Model

      • Field: guardrail.input.model
      • Default: gpt-4o-mini
      • Description: LLM model to use for guardrail validation (OpenAI only)
      • Environment Variable: AK_GUARDRAIL__INPUT__MODEL
    • ID

      • Field: guardrail.input.id
      • Default: None
      • Description: AWS Bedrock guardrail ID (Bedrock only)
      • Environment Variable: AK_GUARDRAIL__INPUT__ID
    • Version

      • Field: guardrail.input.version
      • Default: DRAFT
      • Description: AWS Bedrock guardrail version (Bedrock only)
      • Environment Variable: AK_GUARDRAIL__INPUT__VERSION
  • Output Guardrails

    • Enabled

      • Field: guardrail.output.enabled
      • Default: false
      • Description: Enable output validation guardrails
      • Environment Variable: AK_GUARDRAIL__OUTPUT__ENABLED
    • Type

      • Field: guardrail.output.type
      • Default: openai
      • Options: openai, bedrock, walledai
      • Description: Guardrail provider type
      • Environment Variable: AK_GUARDRAIL__OUTPUT__TYPE
    • Config Path

      • Field: guardrail.output.config_path
      • Default: None
      • Description: Path to guardrail configuration JSON file (OpenAI only)
      • Environment Variable: AK_GUARDRAIL__OUTPUT__CONFIG_PATH
    • Model

      • Field: guardrail.output.model
      • Default: gpt-4o-mini
      • Description: LLM model to use for guardrail validation (OpenAI only)
      • Environment Variable: AK_GUARDRAIL__OUTPUT__MODEL
    • ID

      • Field: guardrail.output.id
      • Default: None
      • Description: AWS Bedrock guardrail ID (Bedrock only)
      • Environment Variable: AK_GUARDRAIL__OUTPUT__ID
    • Version

      • Field: guardrail.output.version
      • Default: DRAFT
      • Description: AWS Bedrock guardrail version (Bedrock only)
      • Environment Variable: AK_GUARDRAIL__OUTPUT__VERSION

Guardrail Setup:

To use OpenAI guardrails, install the openai-guardrails package:

uv add "agentkernel[openai]"

To use AWS Bedrock guardrails, install the AWS package:

uv add "agentkernel[aws]"

To use Walled AI guardrails, install the Walled AI package:

uv add "agentkernel[walledai]"

Create guardrail configuration:

For OpenAI: Create configuration files following the OpenAI Guardrails format.

For Bedrock: Create a guardrail in AWS Bedrock and note the guardrail ID and version.

For Walled AI: Set WALLED_API_KEY, use guardrail type walledai, and control PII masking with pii.

Configure guardrails in your configuration:

OpenAI Example:

guardrail:
  input:
    enabled: true
    type: openai
    model: gpt-4o-mini
    config_path: /path/to/guardrails_input.json
  output:
    enabled: true
    type: openai
    model: gpt-4o-mini
    config_path: /path/to/guardrails_output.json

Bedrock Example:

guardrail:
  input:
    enabled: true
    type: bedrock
    id: your-guardrail-id
    version: "1"  # or "DRAFT"
  output:
    enabled: true
    type: bedrock
    id: your-guardrail-id
    version: "1"

Walled AI Example:

guardrail:
  input:
    enabled: true
    type: walledai
    pii: true
  output:
    enabled: true
    type: walledai
    pii: true

Sandbox Configuration

Enable the sandbox capability to let agents execute code and shell commands in an isolated, permission-bounded environment. When enabled, agents automatically gain sandbox tools (run_code, run_command, write_sandbox_file, read_sandbox_file, check_sandbox_task, list_sandbox_sessions, new_sandbox_session, destroy_sandbox_session, plus install_packages when a profile permits package installs) and the usage guidance is injected into their system prompt.

Minimal single-backend form (a default profile is synthesized from type + its config block):

sandbox:
  enabled: true
  type: local_subprocess        # provider short name, or a dotted path to a SandboxProvider subclass
  local_subprocess: {}
  broker:
    flavor: thread              # thread (local default) | embedded

Full form with explicit workload profiles (provider + lifetime + policy + identity):

sandbox:
  enabled: true
  agents: [coder]               # optional: attach tools/prompt only to these agents; omit = all
  default_profile: workspace
  principal_resolver: null      # optional dotted path to a PrincipalResolver; null = agent identity
  tool_output_max_chars: 8000
  broker:
    flavor: thread
    wait_timeout: 60.0          # seconds before a sync wait promotes to a background task (0 = always)
  profiles:
    workspace:
      type: docker              # container-isolated; needs the sandbox-docker extra + a Docker daemon
      scope: per_session        # per_call | per_session | per_runtime
      idle_timeout: 1800        # seconds of inactivity before the sandbox is reset on next touch
      identity:
        mode: agent             # agent | user
      policy:
        network_egress: deny    # allow | deny | allowlist
        cpu: 1.0
        memory_mb: 512
        timeout: 30.0           # per-execution wall-clock seconds (always enforced)
        strict: true            # fail closed when the provider can't enforce a policy dimension
        languages: [python]     # run_code allowlist; omitted = provider-declared; [] = no run_code
        package_install:        # omitted = installs denied; present = install_packages permitted
          ecosystems: [python]
      docker:
        image: python:3.12-slim

Key fields:

  • enabled (AK_SANDBOX__ENABLED, default false) — master switch; inert when off.
  • agents — agent names the tools/prompt attach to; omit for all agents.
  • type (per profile) — local_subprocess (no isolation; dev/test), docker (container isolation; sandbox-docker extra), kubernetes (pod per sandbox; kubernetes extra), e2b (managed micro-VMs; e2b extra), daytona (cloud containers; daytona extra), ec2_ssm (attach to an existing EC2 instance via SSM; aws extra), or a dotted path to your own SandboxProvider.
  • scope — per_call (fresh per execution), per_session (persists across turns), per_runtime (one shared sandbox per profile).
  • environment — managed (default; the provider creates and disposes sandboxes) or attached (deliberately connect to an existing environment the framework never owns, e.g. an EC2 instance via ec2_ssm; requires the provider's attach_to and is validated against the provider's lifecycle capabilities at startup).
  • policy — network egress, filesystem paths, cpu/memory, timeout; enforced per provider, fail-closed under strict. Also the permissions: languages (which run_code languages are allowed) and package_install (whether install_packages is allowed, and for which ecosystems). A provider's package_installers sets the installer command per ecosystem (default pip install), and on docker/kubernetes/ec2_ssm its language_runtimes sets the interpreter per language (default python -c; on ec2_ssm each argv reads the code from stdin, default python3 -).
  • identity.mode + principal_resolver — run code under the agent's or the invoking user's identity; the default resolver reads the run's Principal, so user mode needs no custom resolver once an identity block authenticates the caller.
  • broker.flavor — thread (default, for CLI/REST), embedded (inline/synchronous), or queue (a separate QueueBrokerWorker fleet over a queue transport).

See the Sandbox guide for the full reference and the examples/sandbox and examples/sandbox/identity examples.

Messaging Platform Integrations

Configure integrations with messaging platforms.

Slack
  • Agent

    • Field: slack.agent
    • Default: ""
    • Description: Default agent for Slack interactions
    • Environment Variable: AK_SLACK__AGENT
  • Agent Acknowledgement

    • Field: slack.agent_acknowledgement
    • Default: ""
    • Description: Acknowledgement message when Slack message is received
    • Environment Variable: AK_SLACK__AGENT_ACKNOWLEDGEMENT
WhatsApp
  • Agent

    • Field: whatsapp.agent
    • Default: ""
    • Description: Default agent for WhatsApp interactions
    • Environment Variable: AK_WHATSAPP__AGENT
  • Verify Token, Access Token, App Secret, Phone Number ID, API Version

    • Environment Variables: AK_WHATSAPP__VERIFY_TOKEN, AK_WHATSAPP__ACCESS_TOKEN, AK_WHATSAPP__APP_SECRET, AK_WHATSAPP__PHONE_NUMBER_ID, AK_WHATSAPP__API_VERSION
Facebook Messenger
  • Agent

    • Field: messenger.agent
    • Default: ""
    • Description: Default agent for Facebook Messenger interactions
    • Environment Variable: AK_MESSENGER__AGENT
  • Verify Token, Access Token, App Secret, API Version

    • Environment Variables: AK_MESSENGER__VERIFY_TOKEN, AK_MESSENGER__ACCESS_TOKEN, AK_MESSENGER__APP_SECRET, AK_MESSENGER__API_VERSION
Instagram
  • Agent

    • Field: instagram.agent
    • Default: ""
    • Description: Default agent for Instagram interactions
    • Environment Variable: AK_INSTAGRAM__AGENT
  • Instagram Account ID, Verify Token, Access Token, App Secret, API Version

    • Environment Variables: AK_INSTAGRAM__INSTAGRAM_ACCOUNT_ID, AK_INSTAGRAM__VERIFY_TOKEN, AK_INSTAGRAM__ACCESS_TOKEN, AK_INSTAGRAM__APP_SECRET, AK_INSTAGRAM__API_VERSION
Telegram
  • Agent

    • Field: telegram.agent
    • Default: ""
    • Description: Default agent for Telegram interactions
    • Environment Variable: AK_TELEGRAM__AGENT
  • Bot Token, Webhook Secret, API Version

    • Environment Variables: AK_TELEGRAM__BOT_TOKEN, AK_TELEGRAM__WEBHOOK_SECRET, AK_TELEGRAM__API_VERSION
Microsoft Teams
  • Agent

    • Field: teams.agent
    • Default: ""
    • Description: Default agent for Microsoft Teams interactions
    • Environment Variable: AK_TEAMS__AGENT
  • Agent Acknowledgement

    • Field: teams.agent_acknowledgement
    • Default: ""
    • Description: Message sent as an acknowledgement when a Teams message is received
    • Environment Variable: AK_TEAMS__AGENT_ACKNOWLEDGEMENT
  • App ID, App Password

    • Description: Azure Bot / Entra ID application (client) ID and client secret. Both are required
    • Environment Variables: AK_TEAMS__APP_ID, AK_TEAMS__APP_PASSWORD
  • Tenant ID

    • Field: teams.tenant_id
    • Default: ""
    • Description: Entra ID tenant that owns the bot's app registration. Required only for a single-tenant registration, whose channel tokens must be issued by its own tenant; leave empty for a multi-tenant bot. Also the fallback tenant for the app-only token used to download attachments whose URL is not pre-authenticated, when the incoming activity carries none
    • Environment Variable: AK_TEAMS__TENANT_ID
Gmail
  • Agent

    • Field: gmail.agent
    • Default: "general"
    • Description: Default agent for Gmail interactions
    • Environment Variable: AK_GMAIL__AGENT
  • Client ID, Client Secret, Token File, Poll Interval, Label Filter

    • Environment Variables: AK_GMAIL__CLIENT_ID, AK_GMAIL__CLIENT_SECRET, AK_GMAIL__TOKEN_FILE, AK_GMAIL__POLL_INTERVAL, AK_GMAIL__LABEL_FILTER

Configuration Examples

Environment Variables

Use the AK_ prefix and underscores for nested fields:

export AK_DEBUG=true
export AK_SESSION__TYPE=redis
export AK_SESSION__REDIS__URL=redis://localhost:6379
export AK_SESSION__REDIS__TTL=604800
export AK_SESSION__REDIS__PREFIX=ak:sessions:
export AK_API__HOST=0.0.0.0
export AK_API__PORT=8000
export AK_A2A__ENABLED=true
export AK_MCP__ENABLED=false
export AK_TRACE__ENABLED=true
export AK_TRACE__TYPE=langfuse  # or openllmetry, logfire, cloudwatch
# For Langfuse:
# export LANGFUSE_PUBLIC_KEY=pk-lf-...
# export LANGFUSE_SECRET_KEY=sk-lf-...
# export LANGFUSE_HOST=https://cloud.langfuse.com
# For OpenLLMetry:
# export TRACELOOP_API_KEY=your-api-key
# For Logfire:
# export LOGFIRE_TOKEN=your-write-token
# For CloudWatch:
# export AWS_REGION=us-east-1
# Test harness (loaded from the separate test-config.yaml — see Test Configuration)
export AK_TEST__MODE=fallback  # Options: score, llm, fallback
export AK_TEST__EVALUATOR=deepeval  # Built-in short name (deepeval, opik or jev), or a dotted path to your own AKEvaluator subclass
export AK_TEST__LLM__MODEL=gpt-4o-mini
export AK_TEST__LLM__PROVIDER=openai
export AK_TEST__LLM__EMBEDDING_MODEL=text-embedding-3-small
# Guardrails configuration
export AK_GUARDRAIL__INPUT__ENABLED=false
export AK_GUARDRAIL__INPUT__TYPE=openai
export AK_GUARDRAIL__INPUT__MODEL=gpt-4o-mini
export AK_GUARDRAIL__INPUT__CONFIG_PATH=/path/to/guardrails_input.json
export AK_GUARDRAIL__OUTPUT__ENABLED=false
export AK_GUARDRAIL__OUTPUT__TYPE=openai
export AK_GUARDRAIL__OUTPUT__MODEL=gpt-4o-mini
export AK_GUARDRAIL__OUTPUT__CONFIG_PATH=/path/to/guardrails_output.json
# Walled AI guardrails
export WALLED_API_KEY=your-walledai-api-key
export AK_GUARDRAIL__INPUT__PII=true
export AK_GUARDRAIL__OUTPUT__PII=true
export AK_DEBUG=true
# Messaging platforms (optional)
export AK_SLACK__AGENT=my-agent
export AK_WHATSAPP__AGENT=my-agent
export AK_MESSENGER__AGENT=my-agent
export AK_INSTAGRAM__AGENT=my-agent
export AK_TELEGRAM__AGENT=my-agent
export AK_GMAIL__AGENT=my-agent
export AK_GMAIL__CLIENT_ID=your-google-client-id
export AK_GMAIL__CLIENT_SECRET=your-google-client-secret

.env File

Create a .env file in your working directory:

AK_DEBUG=false
AK_SESSION__TYPE=redis
AK_SESSION__REDIS__URL=rediss://my-redis:6379
AK_SESSION__REDIS__TTL=1209600
AK_SESSION__REDIS__PREFIX=ak:prod:sessions:
AK_API__HOST=0.0.0.0
AK_API__PORT=8080
AK_A2A__ENABLED=true
AK_A2A__URL=http://localhost:8080/a2a
AK_TRACE__ENABLED=true
AK_TRACE__TYPE=langfuse  # or openllmetry, logfire, cloudwatch
# Langfuse credentials (if using langfuse):
# LANGFUSE_PUBLIC_KEY=pk-lf-...
# LANGFUSE_SECRET_KEY=sk-lf-...
# LANGFUSE_HOST=https://cloud.langfuse.com
# OpenLLMetry credentials (if using openllmetry):
# TRACELOOP_API_KEY=your-api-key
# Logfire credentials (if using logfire):
# LOGFIRE_TOKEN=your-write-token
# CloudWatch region (if using cloudwatch; credentials come from the AWS credential chain):
# AWS_REGION=us-east-1

config.yaml

session:
  type: redis
  redis:
    url: redis://localhost:6379
    ttl: 604800
    prefix: "ak:sessions:"
thread: # optional; configures Conversation Thread Support (enabled by mounting AgentThreadRequestHandler)
  type: redis
  redis:
    url: redis://localhost:6379
    ttl: 2592000
    prefix: "ak:thread:"
execution:
  mode: rest_sync
  queues:
    type: sqs # in_memory | sqs | kafka | nats, or a dotted path to a QueueTransport subclass — mandatory whenever this block is declared
    input:
      url: https://queue.example.com/<accountno>/<queuename> # sqs transport only
      max_receive_count: 3
      no_of_consumers: 5 # in-process pipeline + containerized deployments, ignored by serverless deployments
    output:
      url: https://queue.example.com/<accountno>/<queuename> # sqs transport only
      max_receive_count: 3
      no_of_consumers: 5 # in-process pipeline + containerized deployments, ignored by serverless deployments
    # in_memory: {ack_wait: 300.0, dedup_window: 300.0}  # in_memory transport settings, unused otherwise
    # kafka:  # kafka transport settings, unused otherwise
    #   bootstrap_servers: localhost:9092
    #   input_topic: agent-input
    #   output_topic: agent-output
    #   group_id: agent-kernel
    #   dlq_suffix: .dlq
    #   retry_backoff: 2.0
    #   delivery_timeout: 30.0
    #   metadata_timeout: 5.0
    #   client_config: {} # passthrough confluent-kafka producer/consumer settings (SASL, TLS, tuning)
    # nats:  # nats transport settings, unused otherwise
    #   url: nats://localhost:4222
    #   input_stream: AGENT_REQUESTS
    #   input_subject_prefix: chat.req
    #   output_stream: AGENT_REPLIES
    #   output_subject_prefix: chat.out
    #   partitions: 32
    #   ack_wait: 300.0
    #   retry_backoff: 2.0
    #   duplicate_window: 300.0
    #   max_age: 86400.0
    #   request_timeout: 10.0
    #   auto_provision: false # true for local/dev; leave false where NACK CRs own the objects
    # batch_size is set by the deployment tooling — set via AK_EXECUTION__QUEUES__BATCH_SIZE, never here
  response_store:
    type: redis # in_memory | redis | valkey | dynamodb | dotted path — omit for the built-in in_memory store
    retry_count: 5
    delay: 5
    redis: # if this is given, then valkey/dynamodb response store parts cannot be given
      url: redis://localhost:6379
      prefix: "ak:responses:"
      ttl: 3600
    valkey: # if this is given, then redis/dynamodb response store parts cannot be given (requires the `valkey` extra)
      url: valkey://localhost:6379
      prefix: "ak:responses:"
      ttl: 3600
    dynamodb: # if this is given, then redis/valkey response store parts cannot be given
      table_name: table-name
      table_arn: table-arn
      ttl: 3600
api:
  host: 0.0.0.0
  port: 8000
  enabled_routes:
    agents: true
a2a:
  enabled: true
  agents: ["*"]
  url: http://localhost:8000/a2a
  task_store_type: in_memory
mcp:
  enabled: false
  expose_agents: false
  agents: ["*"]
trace:
  enabled: true
  type: langfuse
# Note: test configuration is no longer set here — it lives in a separate
# test-config.yaml file (see the Test Configuration section)
guardrail:
  input:
    enabled: false
    type: openai
    pii: true
    model: gpt-4o-mini
    config_path: /path/to/guardrails_input.json
  output:
    enabled: false
    type: openai
    pii: true
    model: gpt-4o-mini
    config_path: /path/to/guardrails_output.json
  # For Walled AI, set type: walledai, WALLED_API_KEY,
  # and optionally use input/output pii (default: true) to enable/disable PII masking.
slack:
  agent: my-agent
  agent_acknowledgement: "Processing your request..."
whatsapp:
  agent: my-agent
  agent_acknowledgement: "Processing..."
messenger:
  agent: my-agent
instagram:
  agent: my-agent
telegram:
  agent: my-agent
teams:
  agent: my-agent
  app_id: "<azure-app-client-id>"
  app_password: "<azure-app-client-secret>"
gmail:
  agent: my-agent
  poll_interval: 30
  label_filter: "INBOX"

config.json

{
  "debug": false,
  "session": {
    "type": "redis",
    "redis": {
      "url": "redis://localhost:6379",
      "ttl": 604800,
      "prefix": "ak:sessions:"
    }
  },
  "api": {
    "host": "0.0.0.0",
    "port": 8000,
    "enabled_routes": {
      "agents": true
    }
  },
  "a2a": {
    "enabled": true,
    "agents": ["*"],
    "url": "http://localhost:8000/a2a",
    "task_store_type": "in_memory"
  },
  "mcp": {
    "enabled": false,
    "expose_agents": false,
    "agents": ["*"]
  },
  "trace": {
    "enabled": true,
    "type": "langfuse"
  },
  "guardrail": {
    "input": {
      "enabled": false,
      "type": "openai",
      "model": "gpt-4o-mini",
      "config_path": "/path/to/guardrails_input.json"
    },
    "output": {
      "enabled": false,
      "type": "openai",
      "model": "gpt-4o-mini",
      "config_path": "/path/to/guardrails_output.json"
    }
  },
  "slack": {
    "agent": "my-agent",
    "agent_acknowledgement": "Processing your request..."
  },
  "whatsapp": {
    "agent": "my-agent",
    "agent_acknowledgement": "Processing..."
  },
  "messenger": {
    "agent": "my-agent"
  },
  "instagram": {
    "agent": "my-agent"
  },
  "telegram": {
    "agent": "my-agent"
  },
  "teams": {
    "agent": "my-agent",
    "app_id": "<azure-app-client-id>",
    "app_password": "<azure-app-client-secret>"
  },
  "gmail": {
    "agent": "my-agent",
    "poll_interval": 30,
    "label_filter": "INBOX"
  }
}

Configuration Notes

  • Empty environment variables are ignored
  • Unknown fields in files or environment variables are ignored
  • Environment variables override configuration file values
  • Configuration file values override built-in defaults
  • Nested fields use underscore (_) delimiter in environment variables

Test Configuration (test-config.yaml)

Test harness configuration (comparison mode, evaluator backend, llm models) is separate from the application configuration. It is not part of config.yaml — it lives in its own test-config.yaml file, resolved from the current working directory, and is only loaded when the testing utilities (agentkernel.test) are used. A legacy test: section in config.yaml is ignored. See Test Configuration under Configuration Options for the full list of fields and defaults.

The evaluator field accepts three built-in short names — deepeval (default, uv add "agentkernel[test]") or opik (uv add "agentkernel[opik]", Opik by Comet, runs entirely locally with OPIK_TRACK_DISABLE set so no Opik Cloud account or API key is needed) or jev (uv add "agentkernel[jev]", hosted TypeSafe JEV judge; mode: llm only, needs TYPESAFE_API_KEY, and sends the comparison text to api.typesafe.ai) — or a dotted path to your own AKEvaluator subclass.

test-config.yaml:

mode: fallback
evaluator: deepeval
llm:
  model: gpt-4o-mini
  provider: openai
  embedding_model: text-embedding-3-small

Note that the file is un-nested — there is no top-level test: key. If the file is missing, defaults apply silently (score and fallback tests need no configuration file at all).

Override the test config file path:

export AK_TEST_CONFIG_PATH_OVERRIDE=/path/to/test-config.yaml

Environment variables use the AK_TEST__ prefix and override test-config.yaml values:

export AK_TEST__MODE=fallback  # Options: score, llm, fallback
export AK_TEST__EVALUATOR=deepeval
export AK_TEST__LLM__MODEL=gpt-4o-mini
export AK_TEST__LLM__PROVIDER=openai
export AK_TEST__LLM__EMBEDDING_MODEL=text-embedding-3-small

Extensibility

Custom Framework Adapters

To add support for a new framework:

  1. Implement a Runner class for your framework
  2. Create an Agent wrapper class
  3. Create a Module class that registers agents with the Runtime

Example structure:

from agentkernel.core import Agent, Runner, Module

class MyFrameworkRunner(Runner):
    def run(self, agent, prompt, session):
        # Implement framework-specific execution
        pass

class MyFrameworkAgent(Agent):
    def __init__(self, native_agent):
        self.native_agent = native_agent
        self.runner = MyFrameworkRunner()

class MyFrameworkModule(Module):
    def __init__(self, agents):
        super().__init__()
        for agent in agents:
            wrapped = MyFrameworkAgent(agent)
            self.register(wrapped)

Session Management

Sessions maintain state across agent interactions. Framework adapters manage their own session storage within the Session object using namespaced keys:

  • "crewai" — CrewAI session data
  • "langgraph" — LangGraph session data
  • "openai" — OpenAI Agents SDK session data
  • "adk" — Google ADK session data
  • "pydanticai" — Pydantic AI session data (message history)
  • "maf" — Microsoft Agent Framework session data (AgentSession snapshot)
  • "claude" — Claude Agent SDK session data (each agent's Claude session id and transcript)

Access the session in your runner:

def run(self, agent, prompt, session):
    # Get framework-specific data
    my_data = session.get("my_framework", {})
    
    # Process and update data
    my_data["last_prompt"] = prompt
    
    # Update session
    session.set("my_framework", my_data)

Development

Requirements:

  • Python 3.12+
  • uv 0.8.0+

Setup:

git clone https://github.com/yaalalabs/agent-kernel.git
cd agent-kernel/ak-py
uv sync

Run Tests:

uv run pytest
# or: pytest

Code Quality:

The project uses:

  • black — Code formatting
  • isort — Import sorting
  • mypy — Type checking

License

Unless otherwise specified, all content, including all source code files and documentation files in this repository are:

Copyright (c) 2025-2026 Yaala Labs.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

SPDX-License-Identifier: Apache-2.0

Support

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Metadata

Release files for agentkernel 0.10.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for agentkernel 0.10.0
File Size Uploaded Provenance Actions
agentkernel-0.10.0.tar.gz 831.6 kB With provenance Details for agentkernel-0.10.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for agentkernel 0.10.0
File Interpreter ABI Platform Actions
agentkernel-0.10.0-py3-none-any.whl Python 3 none any Details for agentkernel-0.10.0-py3-none-any.whl

Total release size: 1.9 MB

Release files / agentkernel-0.10.0.tar.gz

Download URL agentkernel-0.10.0.tar.gz
Size 831.6 kB
Tags Source
SHA-256 checksum
How to use checksums
02dcf7e652b1ad78752073f25780f5665b876aaff3126c4b9b66678cb8c5db77
BLAKE2b-256 checksum
How to use checksums
67eaffc3896811d10c7acfc525bf0da2c8ccdf71067d8b85356d7bb95c6e9410
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 13, 2026.

Transparency log

Release files / agentkernel-0.10.0-py3-none-any.whl

Download URL agentkernel-0.10.0-py3-none-any.whl
Size 1.1 MB
Tags Python 3
SHA-256 checksum
How to use checksums
c022dc9d0dbb18757a5b41cfd92641b387aca9dfadafbf552de70e6ead1d9c8d
BLAKE2b-256 checksum
How to use checksums
cfba738072d648c54e037161c9ad032667cab34175cf522f1b3af6f19dcef9cb
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 13, 2026.

Transparency log

Security for agentkernel 0.10.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for agentkernel-0.10.0.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 13, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.10.0 This release

2 release files Full provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page