agentlings 0.9.0

Lightweight A2A + MCP single-process agent framework

Agentlings

Lightweight single-process agent framework exposing both A2A and MCP on a single HTTP port.

---

Each agentling is a small, focused AI agent whose identity is defined by a YAML config file — name, description, system prompt, tools, and skills. The framework handles protocol compliance, conversation journaling, and context management. The LLM is the agent; the framework records and replays.

Install

pip install agentlings
# or, isolated in a managed venv:
uv tool install agentlings

Quick start

# Scaffold a new agent in ./my-agent/
agentling init my-agent
cd my-agent

# Add your Anthropic key (or leave blank to point at Ollama via ANTHROPIC_BASE_URL)
$EDITOR .env

# Run from the agent dir
agentling run

agentling init produces a self-contained directory:

my-agent/
├── agent.yaml           # identity, system prompt, tools, sleep config
├── .env                 # AGENT_API_KEY auto-generated; ANTHROPIC_API_KEY blank for you
├── .env.example         # checked into source control as a template
├── .framework-version   # the framework version that scaffolded this dir
├── data/                # journals, memory, conversations
│   └── .migrations      # applied-migrations log
├── skills/              # drop SKILL.md bundles here; uncomment AGENT_SKILLS_DIR to enable
└── tools/               # drop @tool-decorated .py files here; uncomment AGENT_TOOLS_DIR to enable

agentling run reads agent.yaml, .env, and data/ from the current directory. To operate on a different dir without cd-ing in: agentling run --dir /path/to/agent.

Bumping the framework version preserves your data — pip install --upgrade agentlings && agentling upgrade runs any pending data migrations against data/ without touching agent.yaml or .env.

The running agent serves:

• GET /.well-known/agent-card.json — A2A Agent Card (public, no auth)
• POST /a2a — A2A JSON-RPC endpoint
• POST /mcp — MCP Streamable HTTP endpoint

Both protocols are task-aware. Each request becomes a task; the HTTP handler awaits up to AGENT_TASK_AWAIT_SECONDS (default 60) and either returns the final answer inline or yields a task handle the caller polls. A2A clients can opt out of the wait per-request via configuration.return_immediately = true on message/send (A2A v1.0 JSON name returnImmediately) — the handler then enqueues a Task object immediately and the caller polls via tasks/get.

CLI

Command	Purpose
agentling init <name>	Scaffold a new agent directory from a bundled template
agentling run [--dir]	Run the agent server from CWD or the given directory
agentling upgrade [--dir]	Apply pending data migrations after upgrading the framework
agentling memory show	Print the long-term memory store for the agent in CWD
agentling sleep [--date]	Run a one-off sleep cycle
agentling list-tools	List available tools and groups

Running as a daemon

The framework deliberately stays out of service-management — agentling run is just a long-running foreground process that reads agent.yaml, .env, and data/ from its working directory. Wire it into whatever supervisor you already use.

systemd (Linux)

Set up the agent dir once:

sudo useradd -r -s /bin/false agentling
sudo mkdir -p /opt/agentling && sudo chown agentling: /opt/agentling
sudo -u agentling python3 -m venv /opt/agentling/venv
sudo -u agentling /opt/agentling/venv/bin/pip install agentlings
sudo -u agentling /opt/agentling/venv/bin/agentling init . --dir /opt/agentling --force
sudo $EDITOR /opt/agentling/.env   # add ANTHROPIC_API_KEY

Create /etc/systemd/system/agentling.service:

[Unit]
Description=Agentling
After=network.target

[Service]
Type=simple
User=agentling
WorkingDirectory=/opt/agentling
ExecStart=/opt/agentling/venv/bin/agentling run
Restart=on-failure
RestartSec=5

[Install]
WantedBy=multi-user.target

sudo systemctl daemon-reload
sudo systemctl enable --now agentling
sudo journalctl -u agentling -f

To upgrade: sudo -u agentling /opt/agentling/venv/bin/pip install --upgrade agentlings && sudo -u agentling /opt/agentling/venv/bin/agentling upgrade --dir /opt/agentling && sudo systemctl restart agentling.

launchd (macOS)

Set up the agent dir once with agentling init ~/.agentlings/my-agent, then create ~/Library/LaunchAgents/com.donkeywork.agentling.plist:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key>
    <string>com.donkeywork.agentling</string>
    <key>ProgramArguments</key>
    <array>
        <string>/path/to/venv/bin/agentling</string>
        <string>run</string>
    </array>
    <key>WorkingDirectory</key>
    <string>/Users/you/.agentlings/my-agent</string>
    <key>KeepAlive</key>
    <true/>
    <key>StandardErrorPath</key>
    <string>/tmp/agentling.err</string>
</dict>
</plist>

launchctl load ~/Library/LaunchAgents/com.donkeywork.agentling.plist
tail -f /tmp/agentling.err

Agent definition

Agent identity lives in a YAML file (agent.yaml):

name: k3s-agentling
description: A k3s cluster management agent

tools:
  - bash
  - filesystem

skills:
  - id: k8s-ops
    name: Kubernetes Operations
    description: Manage cluster resources, diagnose issues, apply manifests
    tags: [kubernetes, k3s, devops]
  - id: file-management
    name: File Management
    description: Read, write, and search configuration files
    tags: [files, yaml]

system_prompt: |
  You are a DevOps engineer managing a k3s Kubernetes cluster.

  All configuration changes go through /mnt/lab/k3s as the source of truth.
  Never use kubectl patch/edit/set directly — write manifests and apply them.

  Before any destructive operation, describe the impact and ask for confirmation.

Point to it with AGENT_CONFIG=./agent.yaml.

Available tools

Group	Tools	Description
bash	bash	Shell command execution with timeout
filesystem	read_file, write_file, edit_file, list_directory, search_files	File operations with offset/limit, find-and-replace, glob search
memory	memory_edit	Read and write the agent's persistent long-term memory

Tools are off by default. Run agentling list-tools for details.

Custom tools

Beyond the built-ins, you can author your own tools as plain typed Python functions. Decorate them with @tool, drop the file in a directory, and point AGENT_TOOLS_DIR at it — the agentling scans the directory at startup and registers every Tool it finds.

# tools/weather.py
import os
from typing import Annotated, Literal
from pydantic import Field
from agentlings.tools import tool


@tool
async def weather(
    city: Annotated[str, Field(description="City name, e.g. 'Dublin'.")],
    units: Literal["metric", "imperial"] = "metric",
) -> str:
    """Look up current weather for a city."""
    api_key = os.environ["WEATHER_API_KEY"]
    # ...fetch and return a string the LLM can read...

Then run with AGENT_TOOLS_DIR=./tools agentling run. No registration step, no schema dict — the JSON Schema the LLM sees is derived from the function signature via Pydantic.

How discovery works

• The loader scans the top level of AGENT_TOOLS_DIR for .py files (no recursion).
• Files whose name begins with _ are skipped (use them for shared helpers).
• Each file is imported in isolation — the directory is never added to sys.path, so a file named json.py cannot shadow the stdlib.
• Every module-level Tool instance (i.e. anything you decorated with @tool) is registered.
• An import or registration failure on one file is logged and the scan continues — one broken tool cannot brick the agent.

Authoring contract

Concept	How to express it
Tool name	func.__name__ (or @tool(name="..."))
Tool description	The function's docstring (or @tool(description="..."))
Parameter description / constraints	Annotated[T, Field(description="...", ge=..., le=...)]
Allowed values	Literal["a", "b"] or a str/int Enum
Optional / defaults	A normal Python default (x: int = 30)
Async I/O	async def — sync functions are fine too; both are awaited uniformly
Per-tool secrets	Read your own env vars inside the function (the framework stays out of secret plumbing)

Untyped parameters, *args, **kwargs, and positional-only parameters are rejected at decoration time — @tool raises ToolDefinitionError so misuse fails loudly at startup, not in production.

Reference tools showcasing each pattern live in agentlings.tools.examples (echo, http_get, set_severity, geocode).

Skills

Skills are bundled instructions the agent activates on demand. Each skill is a directory containing a SKILL.md whose YAML frontmatter (name, description) is loaded into the system prompt at startup; the body — and any sibling scripts/, references/, or assets/ — stays on disk until the agent decides the task needs it. This is the progressive disclosure model from the Open Skills specification: metadata is cheap, instructions are loaded on activation, resources are loaded on demand.

skills/
├── pdf-processing/
│   ├── SKILL.md
│   ├── scripts/extract.py
│   └── references/FORMS.md
└── data-analysis/
    └── SKILL.md

A minimal SKILL.md:

---
name: pdf-processing
description: Extract text and tables from PDFs, fill PDF forms, merge files. Use when the user mentions PDFs, forms, or document extraction.
---

Step-by-step instructions for the agent go below the frontmatter.
Reference companion files with relative paths, e.g. `scripts/extract.py`.

Skills are opt-in: set AGENT_SKILLS_DIR=./skills (or any path) and drop skill directories there. On startup the agentling discovers them and prepends a single block to the system prompt explaining progressive disclosure and listing each skill's name, absolute path, and description. The agent reads SKILL.md itself when a task calls for the skill.

Frontmatter constraints

Per the Open Skills spec:

Field	Required	Constraint
name	Yes	1–64 chars, lowercase a-z, digits, hyphens; no leading/trailing/consecutive hyphens; must match the parent directory name
description	Yes	1–1024 chars, non-empty

Optional fields (license, compatibility, metadata, allowed-tools) are accepted but currently ignored at the runtime layer. Malformed skills (missing fields, invalid names, broken YAML) are logged at WARNING and skipped — one bad skill does not prevent the agent from booting.

Discovery is strictly read-only — the agentling never writes to, deletes from, or modifies anything under AGENT_SKILLS_DIR. AGENT_SKILLS_DIR and AGENT_TOOLS_DIR share the same opt-in semantics: unset means "don't scan." AGENT_TOOLS_DIR additionally never adds the user-tools directory to sys.path, so a file named json.py cannot shadow the stdlib.

Naming note: the skills: array in agent.yaml is unrelated — those are A2A Agent Card capabilities advertised on the wire. Runtime skills (this section) live on disk under AGENT_SKILLS_DIR.

Docker

The simplest containerised setup uses the same init + run flow:

FROM python:3.12-slim
WORKDIR /agent
RUN pip install agentlings
RUN agentling init . --api-key dev
VOLUME ["/agent/data"]
EXPOSE 8420
CMD ["agentling", "run"]

docker build -t agentling:latest .
docker run -e ANTHROPIC_API_KEY=sk-ant-... -p 8420:8420 -v ./data:/agent/data agentling:latest

For production, mount your own agent.yaml and .env over the scaffolded ones (or skip the agentling init build step entirely and bind-mount a host directory with everything pre-populated).

Environment variables

Secrets and runtime settings stay in env vars or, more commonly, the .env file inside the agent directory. agentling init creates an .env with AGENT_API_KEY already populated; everything else is opt-in.

Variable	Default	Description
AGENT_CONFIG	./agent.yaml (when present)	Path to agent YAML definition
ANTHROPIC_API_KEY	—	Anthropic API key (required for api.anthropic.com; optional with ANTHROPIC_BASE_URL pointed at e.g. Ollama)
ANTHROPIC_BASE_URL	—	Override the Messages endpoint. Use http://localhost:11434 to target Ollama's Anthropic-compatible API
AGENT_API_KEY	—	API key for authenticating clients
AGENT_MODEL	claude-sonnet-4-6	Model ID — set to an Ollama model (e.g. qwen3-coder) when using ANTHROPIC_BASE_URL
AGENT_MAX_TOKENS	4096	Max tokens per LLM response
AGENT_HOST	0.0.0.0	Bind address
AGENT_PORT	8420	Bind port
AGENT_DATA_DIR	./data	JSONL journal storage directory
AGENT_TOOLS_DIR	—	Directory of @tool-decorated .py files to load at startup
AGENT_SKILLS_DIR	—	Directory of Open Skills SKILL.md bundles to advertise to the agent
AGENT_TASK_AWAIT_SECONDS	60	How long the HTTP handler blocks for task completion before returning a working task handle
AGENT_LOG_LEVEL	INFO	Log level
AGENT_LLM_BACKEND	anthropic	anthropic or mock
AGENT_EXTERNAL_URL	—	Public URL for Agent Card (needed in Docker/k8s)
AGENT_OTEL_ENDPOINT	—	OpenTelemetry collector endpoint
AGENT_OTEL_PROTOCOL	http	Collector protocol (http or grpc)
AGENT_OTEL_INSECURE	true	Disable TLS for collector connection
AGENT_OTEL_HEADERS	—	Comma-separated key=value pairs for collector auth

Memory

Agentlings can maintain persistent long-term memory — a curated set of key-value facts that survive across conversations. Memory transforms an agent from a tool that forgets into one that learns.

How it works

Memory is a JSON file (data/memory/memory.json) containing entries like:

{
  "entries": [
    {
      "key": "cluster-node-count",
      "value": "4 nodes: node1 (control), node2-4 (workers)",
      "recorded": "2026-04-01T10:00:00Z"
    }
  ]
}

The memory block is injected into the system prompt on every LLM call, between the agent's identity and the conversation history. The agent sees its accumulated knowledge as working context, not as a separate tool call.

The memory tool

When the memory tool group is enabled, the agent gets a memory_edit tool with three operations:

Operation	Description
set	Upsert an entry by key. Updates the timestamp.
remove	Delete an entry by key.
list	Return all current entries.

The agent decides what to remember based on its system prompt. A DevOps agent might store cluster topology and known issues. A support agent might store escalation paths and recurring problems.

CLI

# Show current memory
agentling memory show

Configuration

memory:
  token_budget: 2000        # max tokens for the memory block in the system prompt
  # injection_prompt: null   # override the memory/data-dir-awareness template

Sleep cycle

The sleep cycle is a nightly process that journals the day's activity, consolidates new knowledge into memory, prunes stale entries, and cleans up old files. It maps to biological sleep phases.

graph LR
    L[Light Sleep<br/>Gate check] --> D[Deep Sleep<br/>Replay & journal]
    D --> R[REM<br/>Integrate & prune]
    R --> H[Housekeeping<br/>Retention cleanup]

Phase 1: Light sleep — gate check

Quick check: were there any conversations today? If not, skip everything. No LLM calls, no cost.

Phase 2: Deep sleep — replay and file

For each conversation from today, the sleep cycle reads the JSONL journal from the last compaction marker and submits all summaries as a single batch request to the Anthropic Message Batches API. Batch processing runs at 50% cost and processes in parallel.

Each summary call receives the agent's system prompt (so the agent's persona shapes what it considers important), current memory, and the conversation content. The LLM returns a structured ConversationSummary with a narrative and memory candidates.

Results are written to data/journals/YYYY-MM-DD.md.

Phase 3: REM — integrate and prune

A single LLM call receives current memory, today's journal, and all extracted memory candidates. It integrates new facts, deduplicates, reviews existing entries for staleness, and returns a ConsolidatedMemory — the complete updated memory store. Written atomically to memory.json.

Phase 4: Housekeeping — retention cleanup

Deletes conversation JSONL files older than conversation_retention_days and journal files older than journal_retention_days.

Configuration

sleep:
  schedule: "0 2 * * *"           # cron expression (default: 2am daily)
  journal_retention_days: 30       # keep journals for 30 days
  conversation_retention_days: 14  # keep JSONL conversations for 14 days
  memory_max_entries: 50           # hard cap after consolidation
  # model: null                    # override model for sleep calls
  # summary_prompt: null           # override per-conversation summary prompt
  # consolidation_prompt: null     # override REM consolidation prompt

CLI

# Trigger sleep cycle manually
agentling sleep --date 2026-04-01

Data directory layout

data/
  abc123.jsonl           # conversation journals (flat, as before)
  def456.jsonl
  memory/
    memory.json          # persistent memory store
  journals/
    2026-04-01.md        # daily sleep journal
    2026-04-02.md

The agent is told about this directory structure and can use its filesystem tools to search past journals and conversation logs for context beyond what fits in memory.

OpenTelemetry

Off by default. When enabled, the framework emits a comprehensive trace + metric surface over OTLP — every HTTP request, every protocol call, every task, every LLM completion, every tool, and every journal write.

telemetry:
  enabled: true
  endpoint: "http://otel-collector:4318"
  protocol: "http"                        # "http" or "grpc"
  service_name: "agentling"
  insecure: true
  headers:                                # optional auth headers
    Authorization: "Bearer your-token"

Or via env vars (these override the YAML and force enabled: true):

AGENT_OTEL_ENDPOINT=http://collector:4318
AGENT_OTEL_PROTOCOL=http      # or grpc
AGENT_OTEL_INSECURE=true
AGENT_OTEL_HEADERS="Authorization=Bearer tok,X-Tenant=team-a"

Spans emitted

A request flows top-to-bottom through this tree, parent → child:

Span	Where
agentling.http.request	Starlette middleware. Extracts inbound traceparent so external traces stitch in.
agentling.a2a.execute / agentling.a2a.cancel	A2A executor.
agentling.mcp.call_tool	MCP server.
agentling.engine.spawn / .poll / .cancel	Task engine entry points.
agentling.task.worker	Per-task worker. Stamps token totals as attributes.
agentling.completion	LLM completion cycle. Stamps cycle-wide token totals.
agentling.completion.llm_call	One LLM turn. Per-turn token attributes.
agentling.completion.tool_exec	Per-tool execution within a turn.
agentling.llm.complete / .count_tokens / .batch_create / .batch_status / .batch_results	LLM client calls. Token usage attributes on the complete span.
agentling.engine.recovery	Startup crash-recovery pass.
agentling.sleep.*	Nightly sleep cycle phases (when enabled).
agentling.memory.list/set/remove	Memory tool operations.

Metrics emitted

Tokens (per-call histograms + monotonic counters, labeled by llm.model, llm.path of live/batch):

• agentling.llm.input_tokens / _total
• agentling.llm.output_tokens / _total
• agentling.llm.cache_creation_input_tokens / _total
• agentling.llm.cache_read_input_tokens / _total
• agentling.llm.total_tokens
• agentling.llm.cache_hit_ratio
• agentling.llm.calls_total

Tasks (counters + active gauge):

• agentling.tasks.spawned_total / completed_total / failed_total / cancelled_total
• agentling.tasks.active
• agentling.tasks.context_busy_rejections_total
• agentling.tasks.crash_recovery_repaired_total / crash_recovery_failed_total

Completion + tools:

• agentling.completion.duration_seconds, .turns
• agentling.tool.calls, .errors, .duration_seconds (labeled by tool.name, tool.is_error)

Journal I/O:

• agentling.journal.append_seconds (labeled by journal.target=parent|sub)
• agentling.journal.replay_seconds
• agentling.journal.bytes_appended_total
• agentling.journal.entries_replayed_total

HTTP:

• agentling.http.request span carries http.status_code, http.duration_seconds, http.method, http.target.

When telemetry is disabled (the default) or the OpenTelemetry packages are not installed, all instrumentation is a no-op — the cost is one function call per span/metric site.

Extended thinking

Agentlings can be configured to let Claude do extended reasoning before its visible reply. Off by default. Three modes match the model-generation split that landed in Q1 2026:

Mode	When to use	YAML block
off (default)	Non-Anthropic backends; cost-sensitive workloads	thinking: { mode: "off" }
adaptive	Opus 4.6+, Sonnet 4.6+, Mythos Preview	thinking: { mode: "adaptive", effort: "medium" }
budget	Sonnet 3.7 / 4 / 4.5, Opus 4 / 4.1 / 4.5, Haiku 4.5	thinking: { mode: "budget", budget_tokens: 4096, interleaved: true }

Adaptive mode (recommended on 4.6+)

The model picks its own thinking budget per request. You control depth with the effort parameter (low | medium | high | xhigh | max). Interleaved thinking between tool calls is automatic — no flag needed. On Opus 4.7+ thinking content is hidden by default; set display: "summarized" to opt back in.

thinking:
  mode: adaptive
  effort: medium
  # display: summarized  # opt back into summarized thinking on Opus 4.7+

Required on Opus 4.7+. Recommended on Opus 4.6 and Sonnet 4.6 (where legacy budget_tokens is deprecated but still accepted).

Budget mode (legacy)

Sets thinking: {"type": "enabled", "budget_tokens": N} on every Messages call. budget_tokens must be ≥ 1024 and (unless interleaved: true) less than max_tokens. Setting interleaved: true adds the interleaved-thinking-2025-05-14 beta header so the model can think between tool calls; in that case the budget is a per-turn total and may exceed max_tokens.

thinking:
  mode: budget
  budget_tokens: 4096
  interleaved: true

Required on Sonnet 4 / 4.5, Opus 4 / 4.1 / 4.5, and Haiku 4.5 (which does not support interleaved). Rejected with HTTP 400 on Opus 4.7+.

Sleep cycle

The sleep cycle's batch calls inherit the same thinking config. Interleaved thinking is silently dropped for batch calls (the batches API cannot carry a per-request beta header), but the thinking block itself is still sent.

Mock backend

The mock backend ignores thinking for behaviour but records the config on the client (MockLLMClient.thinking_config) so tests can assert the wiring without an Anthropic key.

Architecture

graph TB
    A2A[A2A Client] -->|POST /a2a| A2ASDK[a2a-sdk Server]
    MCP[MCP Client] -->|POST /mcp| MCPSDK[mcp SDK Server]
    A2ASDK --> Executor[AgentlingExecutor]
    Executor --> Loop[MessageLoop]
    MCPSDK --> Loop
    Loop --> Store[JSONL Store]
    Loop --> LLM[LLM Client]
    Loop --> Tools[Tool Registry]

Both protocols feed into a single MessageLoop.process_message() entrance. Conversations are persisted as append-only JSONL journals with compaction markers as replay cursors.

Testing

# Unit tests (no network, no LLM)
pytest tests/unit/ -v

# Integration tests (starts real server with mock LLM)
pytest tests/integration/ -v

# All tests
pytest tests/ -v

Integration tests use native SDK clients — a2a-sdk ClientFactory for A2A and mcp ClientSession for MCP — talking to a real server over HTTP. All LLM responses are mocked.

Built with

• a2a-sdk — A2A protocol server + client
• mcp — MCP protocol server + client
• anthropic — LLM backend
• starlette + uvicorn — HTTP server