agentbreeder 1.7.0

Define Once. Deploy Anywhere. Govern Automatically.

AgentBreeder

Stop wrangling agents. Start shipping them.

One YAML file. Any framework. Any cloud. Governance built in.

Quick Start · How It Works · Install · Features · CLI Reference · Docs · Contributing

---

Your company has 47 AI agents. Nobody knows what they cost, who approved them, or which ones are still running. Three teams built the same summarizer. The security team hasn't audited any of them.

AgentBreeder fixes this.

Write one agent.yaml. Run agentbreeder deploy. Your agent is live — with RBAC, cost tracking, audit trail, and org-wide discoverability. Automatic. Not optional.

╔═══════════════════════════════════════════════════════════════╗
║                   AGENTBREEDER DEPLOY                         ║
╠═══════════════════════════════════════════════════════════════╣
║                                                               ║
║  ✅  YAML parsed & validated                                  ║
║  ✅  RBAC check passed (team: engineering)                    ║
║  ✅  Dependencies resolved (3 tools, 1 prompt)                ║
║  ✅  Container built (langgraph runtime)                      ║
║  ✅  Deployed to GCP Cloud Run                                ║
║  ✅  Health check passed                                      ║
║  ✅  Registered in org registry                               ║
║  ✅  Cost attribution: engineering / $0.12/hr                 ║
║                                                               ║
║  ENDPOINT: https://support-agent-a1b2c3.run.app              ║
║  STATUS:   ✅ LIVE                                            ║
║                                                               ║
╚═══════════════════════════════════════════════════════════════╝

---

The Problem

AI coding tools make it easy to build agents. Nobody has made it easy to ship them responsibly.

What happens today	What happens with AgentBreeder
Every framework has its own deploy story	One YAML, any framework, any cloud
No RBAC — anyone deploys anything	RBAC validated before the first container builds
No cost tracking — $40k surprise cloud bills	Cost attributed per team, per agent, per model
No audit trail — "who deployed that?"	Every deploy logged with who, what, when, where
No discoverability — duplicate agents everywhere	Org-wide registry — search before you build
Governance is bolted on after the fact	Governance is a structural side effect of deploying

Governance is not configuration. It is a side effect of the pipeline. There is no way to skip it.

---

How It Works

# agent.yaml — this is the entire config
name: customer-support-agent
version: 1.0.0
team: customer-success
owner: [e-mail]

framework: langgraph          # or: openai_agents, claude_sdk, crewai, google_adk, custom

model:
  primary: claude-sonnet-4
  fallback: gpt-4o

tools:
  - ref: tools/zendesk-mcp    # pull from org registry
  - ref: tools/order-lookup

deploy:
  cloud: gcp                  # or: aws, local, kubernetes
  scaling:
    min: 1
    max: 10

pip install agentbreeder
agentbreeder deploy ./agent.yaml

That's it. Eight atomic steps — parse, RBAC, resolve deps, build container, provision infra, deploy, health check, register. If any step fails, the entire deploy rolls back.

---

Three Ways to Build

All three tiers compile to the same internal format. Same deploy pipeline. Same governance. No lock-in.

Tier	Who	How	Eject to
No Code	PMs, analysts, citizen builders	Visual drag-and-drop canvas — pick model, tools, prompts from the registry	Low Code (view YAML)
Low Code	ML engineers, DevOps	Write agent.yaml in any IDE	Full Code (agentbreeder eject)
Full Code	Senior engineers, researchers	Python/TS SDK with full programmatic control	—

# Full Code SDK — builder pattern
from agenthub import Agent

agent = (
    Agent("support-agent", version="1.0.0", team="eng")
    .with_model(primary="claude-sonnet-4", fallback="gpt-4o")
    .with_tools(["tools/zendesk-mcp", "tools/order-lookup"])
    .with_prompt(system="You are a helpful customer support agent.")
    .with_deploy(cloud="gcp", min_scale=1, max_scale=10)
)
agent.deploy()

---

What's Implemented

Frameworks

Framework	Status	Runtime
LangGraph	✅	engine/runtimes/langgraph.py
OpenAI Agents SDK	✅	engine/runtimes/openai_agents.py
Claude SDK (Anthropic)	✅	engine/runtimes/claude_sdk.py
CrewAI	✅	engine/runtimes/crewai.py
Google ADK	✅	engine/runtimes/google_adk.py
Custom (bring your own)	✅	engine/runtimes/custom.py

Cloud Targets

Target	Status	Deployer
Local (Docker Compose)	✅	engine/deployers/docker_compose.py
GCP Cloud Run	✅	engine/deployers/gcp_cloudrun.py
AWS ECS Fargate	🔲 Planned
Kubernetes	🔲 Planned

LLM Providers (6 providers + fallback chains)

Provider	Status
Anthropic (Claude)	✅
OpenAI (GPT-4o, o1, etc.)	✅
Google (Gemini)	✅
Ollama (local models)	✅
LiteLLM gateway (100+ models)	✅
OpenRouter	✅

Secrets Backends

Backend	Status
Environment variables / .env	✅
AWS Secrets Manager	✅
GCP Secret Manager	✅
HashiCorp Vault	✅

Platform Features (30+ shipped)

Feature	Status
Org-wide agent registry	✅
Visual agent builder (ReactFlow canvas)	✅
Multi-agent orchestration (6 strategies)	✅
Visual orchestration canvas	✅
A2A (Agent-to-Agent) protocol	✅
MCP server hub + sidecar injection	✅
Agent evaluation framework	✅
Cost tracking (per team / agent / model)	✅
RBAC + team management	✅
Full audit trail	✅
Distributed tracing (OpenTelemetry)	✅
AgentOps fleet dashboard	✅
Community marketplace + templates	✅
Git workflow (PR create → review → publish)	✅
Prompt builder + test panel	✅
RAG index builder	✅
Memory configuration	✅
Tool sandbox execution	✅
Interactive chat playground	✅
API versioning (v1 stable, v2 preview)	✅
Python SDK	✅
TypeScript SDK	✅
Tier mobility (agentbreeder eject)	✅

---

Orchestration

Six strategies. Define in YAML or the visual canvas — both compile to the same pipeline.

# orchestration.yaml
name: support-pipeline
version: "1.0.0"
team: customer-success
strategy: router       # router | sequential | parallel | hierarchical | supervisor | fan_out_fan_in

agents:
  triage:
    ref: agents/triage-agent
    routes:
      - condition: billing
        target: billing
      - condition: default
        target: general
  billing:
    ref: agents/billing-agent
    fallback: general
  general:
    ref: agents/general-agent

shared_state:
  type: session_context
  backend: redis

deploy:
  target: gcp

Or programmatically with the SDK:

from agenthub import Orchestration

pipeline = (
    Orchestration("support-pipeline", strategy="router", team="eng")
    .add_agent("triage",  ref="agents/triage-agent")
    .add_agent("billing", ref="agents/billing-agent")
    .add_agent("general", ref="agents/general-agent")
    .with_route("triage", condition="billing", target="billing")
    .with_route("triage", condition="default", target="general")
    .with_shared_state(state_type="session_context", backend="redis")
)
pipeline.deploy()

---

Install

PyPI (recommended)

# Full CLI + API server + engine
pip install agentbreeder

# Lightweight Python SDK only (for programmatic agent definitions)
pip install agentbreeder-sdk

npm (TypeScript / JavaScript)

npm install @agentbreeder/sdk

import { Agent } from "@agentbreeder/sdk";

const agent = new Agent("customer-support", { version: "1.0.0", team: "eng" })
  .withModel({ primary: "claude-sonnet-4", fallback: "gpt-4o" })
  .withTool({ ref: "tools/zendesk-mcp" })
  .withDeploy({ cloud: "aws", region: "us-east-1" });

await agent.deploy();

Homebrew (macOS / Linux)

brew tap rajitsaha/agentbreeder
brew install agentbreeder

Docker

# API server
docker pull rajits/agentbreeder-api
docker run -p 8000:8000 rajits/agentbreeder-api

# Dashboard
docker pull rajits/agentbreeder-dashboard
docker run -p 3001:3001 rajits/agentbreeder-dashboard

# CLI (for CI/CD pipelines)
docker pull rajits/agentbreeder-cli
docker run rajits/agentbreeder-cli deploy agent.yaml --target gcp

---

Quick Start

pip install agentbreeder

# Scaffold your first agent (interactive wizard — pick framework, cloud, model)
agentbreeder init

# Validate the config
agentbreeder validate agent.yaml

# Deploy locally
agentbreeder deploy agent.yaml --target local

Or run the full platform locally:

git clone https://github.com/rajitsaha/agentbreeder.git
cd agentbreeder

python -m venv venv && source venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env

# Start postgres + redis + API + dashboard
docker compose -f deploy/docker-compose.yml up -d

Dashboard: http://localhost:3001 · API: http://localhost:8000 · API Docs: http://localhost:8000/docs

See docs/quickstart.md for the full guide.

---

CLI

24 commands. Everything you need from scaffold to teardown.

agentbreeder init              # Scaffold a new agent project (interactive wizard)
agentbreeder validate          # Validate agent.yaml without deploying
agentbreeder deploy            # Deploy an agent (the core command)
agentbreeder up / down         # Start / stop the full local platform stack
agentbreeder status            # Show deploy status
agentbreeder logs <name>       # Tail agent logs
agentbreeder list              # List agents / tools / models / prompts
agentbreeder describe <name>   # Show detail for a registry entity
agentbreeder search <query>    # Search across the entire registry
agentbreeder chat <name>       # Interactive chat with a deployed agent
agentbreeder eval              # Run evaluations against golden datasets
agentbreeder eject             # Eject from YAML to Full Code SDK
agentbreeder submit            # Create a PR for review
agentbreeder review            # Review / approve / reject a submission
agentbreeder publish           # Merge approved PR and publish to registry
agentbreeder provider          # Manage LLM provider connections
agentbreeder secret            # Manage secrets across backends (env, AWS, GCP, Vault)
agentbreeder scan              # Discover MCP servers and LiteLLM models
agentbreeder template          # Manage agent templates
agentbreeder orchestration     # Multi-agent orchestration commands
agentbreeder teardown          # Remove a deployed agent and clean up resources

See docs/cli-reference.md for full usage and flags.

---

Architecture

Developer                    AgentBreeder Platform                  Cloud
                                     ┌──────────────┐
agent.yaml  ──▶  [ CLI ]  ──▶  │  API Server  │  ──▶  [ Engine ]  ──▶  AWS / GCP / Local
                                     └──────┬───────┘         │
                                            │                 ▼
                                     ┌──────▼───────┐  ┌─────────────────┐
                                     │  PostgreSQL   │  │ Container Build │
                                     │  (Registry)   │  │  + MCP Sidecar  │
                                     └──────┬───────┘  └─────────────────┘
                                            │
                                     ┌──────▼───────┐
                                     │    Redis      │
                                     │   (Queue)     │
                                     └──────────────┘

Deploy pipeline — 8 atomic steps. If any fails, the entire deploy rolls back:

1. Parse & validate YAML
2. RBAC check (fail fast if unauthorized)
3. Dependency resolution (tools, prompts, models from registry)
4. Container build (framework-specific Dockerfile)
5. Infrastructure provision (Pulumi)
6. Deploy & health check
7. Auto-register in org registry
8. Return endpoint URL

---

Project Structure

agentbreeder/
├── api/                # FastAPI backend — 25 route modules, services, models
├── cli/                # CLI — 24 commands (Typer + Rich)
├── engine/
│   ├── config_parser.py       # YAML parsing + JSON Schema validation
│   ├── builder.py             # 8-step atomic deploy pipeline
│   ├── orchestrator.py        # Multi-agent orchestration engine
│   ├── providers/             # LLM providers (Anthropic, OpenAI, Google, Ollama)
│   ├── runtimes/              # Framework builders (LangGraph, OpenAI Agents)
│   ├── deployers/             # Cloud deployers (Docker Compose, GCP Cloud Run)
│   ├── secrets/               # Secrets backends (env, AWS, GCP, Vault)
│   ├── a2a/                   # Agent-to-Agent protocol
│   └── mcp/                   # MCP server packaging
├── registry/           # Catalog services — agents, tools, models, prompts, templates
├── sdk/python/         # Python SDK (pip install agentbreeder-sdk)
├── sdk/typescript/     # TypeScript SDK (npm install @agentbreeder/sdk)
├── connectors/         # LiteLLM, OpenRouter, MCP scanner
├── dashboard/          # React + TypeScript + Tailwind
├── tests/              # 2,378 unit tests + Playwright E2E
└── examples/           # Working examples per framework + orchestration

---

Documentation

Doc	Description
Quickstart	Local setup in under 10 minutes
CLI Reference	All 24 commands with flags and examples
agent.yaml Reference	Full configuration field reference
orchestration.yaml Reference	Multi-agent pipeline config
Orchestration SDK	Python/TS SDK for complex workflows
API Stability	Versioning and deprecation policy
Local Development	Contributor setup guide
ARCHITECTURE.md	Technical deep-dive
ROADMAP.md	Release plan and milestone status
CONTRIBUTING.md	How to contribute

---

"How is this different from LangGraph / CrewAI / Mastra / OpenAI Agents?"

It's not the same category.

Agent SDKs help you build an agent. AgentBreeder helps you ship, govern, and operate it. You use them together — not instead of each other.

┌─────────────────────────────────────────────────────────────────────┐
│                        What you do today                            │
│                                                                     │
│   LangGraph / CrewAI / Mastra / OpenAI Agents / Claude SDK          │
│   ──────────────────────────────────────────────────────────         │
│   Build the agent            ✅  Great at this                      │
│   Deploy it somewhere        🤷  You figure it out                  │
│   Track who deployed what    🤷  You figure it out                  │
│   Control who can deploy     🤷  You figure it out                  │
│   Track costs per team       🤷  You figure it out                  │
│   Audit every deploy         🤷  You figure it out                  │
│   Discover agents across org 🤷  You figure it out                  │
│   Multi-cloud portability    🤷  You figure it out                  │
└─────────────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────────────┐
│                     What changes with AgentBreeder                  │
│                                                                     │
│   LangGraph / CrewAI / Mastra / OpenAI Agents / Claude SDK          │
│                          +                                          │
│                    AgentBreeder                                     │
│   ──────────────────────────────────────────────────────────         │
│   Build the agent            ✅  Your SDK does this (unchanged)     │
│   Deploy it somewhere        ✅  `agentbreeder deploy` — any cloud  │
│   Track who deployed what    ✅  Automatic audit trail              │
│   Control who can deploy     ✅  RBAC validated before build starts │
│   Track costs per team       ✅  Per team / agent / model           │
│   Audit every deploy         ✅  Every deploy logged                │
│   Discover agents across org ✅  Org-wide searchable registry      │
│   Multi-cloud portability    ✅  Same YAML → AWS, GCP, or local    │
└─────────────────────────────────────────────────────────────────────┘

Think of it this way:

Analogy	Build tool	Ship/Operate tool
Code	You write Python	Docker + Kubernetes deploys it
Infrastructure	Terraform defines it	CI/CD ships it
Agents	LangGraph / CrewAI / Mastra builds it	AgentBreeder ships and governs it

AgentBreeder is to agents what Docker + Kubernetes is to microservices — the deployment, orchestration, and operations layer. Your agent framework is the application; AgentBreeder is the platform.

The real comparison

	Agent SDKs	Agent Platforms	AgentBreeder
	LangGraph, CrewAI, Mastra, OpenAI Agents, Claude SDK, Google ADK	Vertex AI Agent Builder, Amazon Bedrock Agents, Azure AI Agent Service
Purpose	Build agents	Build + deploy (their way)	Ship + govern + operate (any agent, any cloud)
Framework lock-in	✅ You're locked in	✅ You're locked in	❌ Bring any framework
Cloud lock-in	N/A	✅ Their cloud only	❌ AWS, GCP, local, K8s
Works with your existing agents	N/A	❌ Rewrite required	✅ Wrap in agent.yaml, deploy
Governance	❌ Not their problem	⚠️ Partial, platform-specific	✅ RBAC, audit, cost — automatic
Org-wide registry	❌	❌	✅ Every deploy auto-registered
Cost attribution	❌	⚠️ Project-level	✅ Per team / agent / model
Multi-agent orchestration	⚠️ Framework-specific	⚠️ Limited	✅ 6 strategies, cross-framework
MCP native	⚠️ Varies	❌	✅ Hub + sidecar injection
Open source	✅ Most are	❌ All proprietary	✅ Apache 2.0
Vendor lock-in	Framework	Cloud vendor	None

vs. Managed Agent Platforms (detailed)

Every cloud vendor now has an agent platform. They all share the same tradeoff: convenience in exchange for lock-in. AgentBreeder gives you the convenience without the lock-in.

	AWS Bedrock Agents	Vertex AI Agent Builder	Azure AI Agent Service	Databricks AgentBricks	Claude Managed Agents	AgentBreeder
Cloud	AWS only	GCP only	Azure only	Databricks only	Anthropic only	Any cloud + local
Frameworks	Bedrock SDK	Vertex SDK / ADK	Azure SDK	Mosaic Agent Framework	Claude SDK	LangGraph, CrewAI, OpenAI, Claude, ADK, custom
Bring your own agent code	❌ Rewrite in their SDK	❌ Rewrite in their SDK	❌ Rewrite in their SDK	❌ Rewrite in their SDK	❌ Claude only	✅ Any framework, wrap in YAML
Multi-cloud deploy	❌	❌	❌	❌	❌	✅ Same config → any target
Migrate away	Painful (SDK lock-in)	Painful (SDK lock-in)	Painful (SDK lock-in)	Painful (platform lock-in)	Painful (API lock-in)	Trivial (it's your code + YAML)
RBAC	IAM (AWS-specific)	IAM (GCP-specific)	RBAC (Azure-specific)	Unity Catalog	API keys	Framework-agnostic RBAC
Cost tracking	CloudWatch + Cost Explorer	GCP Billing	Azure Cost Management	DBU tracking	Usage dashboard	Built-in: per team / agent / model
Org-wide registry	❌	❌	❌	Unity Catalog (models)	❌	✅ Agents, tools, prompts, models
Multi-agent orchestration	Step Functions (manual)	Limited	Limited	Limited	Tool use (single agent)	6 strategies (router, sequential, parallel, hierarchical, supervisor, fan-out)
MCP support	❌	❌	❌	❌	✅ Native	✅ Hub + sidecar injection
Open source	❌	❌	❌	❌	❌	✅ Apache 2.0
Data residency	Their cloud	Their cloud	Their cloud	Their platform	Their API	Your infrastructure
Pricing	Pay per invocation + hosting	Pay per invocation + hosting	Pay per invocation + hosting	DBU consumption	Pay per token	Free (you pay your own cloud costs)

The pattern: Every managed platform wants you to rewrite your agent in their SDK, deploy to their cloud, and pay their markup. When you want to switch, you rewrite everything.

AgentBreeder's position: Your agent code stays yours. Your cloud stays yours. Your data stays yours. AgentBreeder is the deploy + governance layer that works across all of them.

What AgentBreeder is NOT

• Not a replacement for LangGraph. Use LangGraph to build your agent's logic. Use AgentBreeder to deploy, govern, and operate it.
• Not a replacement for CrewAI. Build your crew in CrewAI. Ship it with AgentBreeder.
• Not a replacement for Mastra. Build your TypeScript agent in Mastra. Wrap it in agent.yaml. Deploy it anywhere.
• Not competing with Bedrock / Vertex / Azure. They're managed services. AgentBreeder is open-source infrastructure. Different model, different tradeoffs.
• Not another agent framework. We don't have opinions about how you build agents. We have opinions about how you ship them responsibly.
• Not a managed service. It's open-source infrastructure you run. No vendor lock-in. No data leaves your cloud.

---

Contributing

High-impact areas where contributions are especially welcome:

• AWS ECS deployer — engine/deployers/aws_ecs.py — most requested cloud target
• Framework runtimes — additional frameworks beyond the six currently supported
• Agent templates — starter templates for common use cases
• Connectors — Datadog, Grafana, and other observability integrations

See CONTRIBUTING.md for setup instructions and guidelines.

---

Community

• GitHub Issues — bugs and feature requests
• GitHub Discussions — questions and show & tell

---

License

Apache License 2.0 — see LICENSE.

---

Built by Rajit Saha

Tech executive · 20+ years building enterprise data & ML platforms · Udemy, LendingClub, VMware, Yahoo

If AgentBreeder saves you time, star the repo and share it with your team.