Low Code — agent.yaml + CLI

Configure agents in YAML and deploy with one command. No infrastructure expertise needed.

Write one agent.yaml file, run agentbreeder deploy. That's it.

Who this is for: ML engineers, DevOps engineers, platform engineers, and architects who are comfortable editing YAML and running terminal commands — but don't want to write boilerplate infrastructure code.

Supported Frameworks

Set framework: to any of the 6 supported values. AgentBreeder picks the right runtime container, installs the correct dependencies, and injects the framework's server adapter automatically.

Value	Framework
`langgraph`	LangGraph (LangChain)
`openai_agents`	OpenAI Agents SDK
`claude_sdk`	Anthropic Claude SDK
`crewai`	CrewAI
`google_adk`	Google Agent Development Kit
`custom`	Bring your own entrypoint

Supported Deployment Targets

Set deploy.cloud and optionally deploy.runtime.

Status: Deploy-target availability (which targets ship full governance, greenfield provisioning, and secret auto-mirror) is maintained in one canonical place — see the deploy-target status table.

`cloud`	`runtime` (default)	Also supports
`local`	`docker-compose`	—
`aws`	`ecs-fargate`	`app-runner`
`gcp`	`cloud-run`	—
`azure`	`container-apps`	—
`kubernetes`	`deployment`	—
`claude-managed`	(no container)	—

Getting Started

1. Install

pipx install agentbreeder    # or: python3 -m pip install agentbreeder

2. Initialize a project

agentbreeder init

The wizard asks 5 questions and generates a ready-to-run project:

my-agent/
├── agent.yaml        # ← the only file AgentBreeder needs
├── agent.py          # working agent code for your chosen framework
├── requirements.txt
├── .env.example
└── README.md

3. Edit agent.yaml

Here is a fully-featured example showing all major sections:

# Identity
name: support-agent
version: 1.0.0
description: "Handles tier-1 customer support"
team: customer-success
owner: alice@company.com
tags: [support, zendesk, production]

# Framework
framework: langgraph   # langgraph | openai_agents | claude_sdk | crewai | google_adk | custom

# Model
model:
  primary: claude-sonnet-4      # any model in your registry
  fallback: gpt-4o              # used if primary is unavailable
  temperature: 0.7
  max_tokens: 4096

# Tools & MCP Servers — always reference by registry ref. Never inline.
# Local files at tools/<snake_name>.py override the registry automatically.
tools:
  - ref: tools/web-search       # resolves engine.tools.standard.web_search
  - ref: tools/markdown-writer  # resolves engine.tools.standard.markdown_writer

# Knowledge Bases (RAG)
knowledge_bases:
  - ref: kb/product-docs
  - ref: kb/return-policy

# Versioned Prompts — always reference by registry ref. Never inline.
# The actual prompt body lives at prompts/<name>.md in the project.
prompts:
  system: prompts/microlearning-system   # file-first; falls back to registry API

# Guardrails
guardrails:
  - pii_detection           # strip PII from outputs
  - hallucination_check     # flag low-confidence responses
  - content_filter          # block harmful content

# Deployment
deploy:
  cloud: aws                # aws | gcp | azure | kubernetes | local | claude-managed
  runtime: app-runner       # default for aws: ecs-fargate; also: app-runner
  region: us-east-1
  scaling:
    min: 1
    max: 10
    target_cpu: 70
  resources:
    cpu: "1"
    memory: "2Gi"
  env_vars:
    LOG_LEVEL: info
  secrets:
    - ZENDESK_API_KEY
    - OPENAI_API_KEY

# Access Control
access:
  visibility: team              # public | team | private
  allowed_callers:
    - team:engineering
    - team:customer-success
  require_approval: false

For the full field reference see agent.yaml Reference →.

Resolver chain

Both prompts: and tools: use a file-first resolution chain so you can iterate locally without touching the registry. AgentBreeder's runtime calls engine.prompt_resolver.resolve_prompt("prompts/<name>", project_root) and engine.tool_resolver.resolve_tool("tools/<kebab>", project_root) at startup.

Prompt: local file (<project>/prompts/<name>.md) → registry API (${AGENTBREEDER_REGISTRY_URL}/api/v1/registry/prompts) → inline literal (backward compat).

Tool: local file (<project>/tools/<snake_name>.py) → first-party stdlib (engine.tools.standard.<snake_name>) → registry API metadata.

TypeScript agents (sdk-ts) get the same precedence via resolvePrompt() and resolveTool() in engine/runtimes/templates/node/_shared_loader.ts, including local .ts/.js/.mjs files.

4. Validate

agentbreeder validate agent.yaml

Runs three checks: YAML syntax, JSON Schema, and semantic validation (framework supported, team exists).

5. Deploy

agentbreeder deploy agent.yaml --target local

agentbreeder deploy agent.yaml --target app-runner --region us-east-1

Required in deploy:

deploy:
  cloud: aws
  runtime: app-runner
  env_vars:
    AWS_ACCOUNT_ID: "123456789012"
    AWS_REGION: us-east-1

agentbreeder deploy agent.yaml --target ecs-fargate --region us-east-1

Required in deploy:

deploy:
  cloud: aws
  runtime: ecs-fargate
  env_vars:
    AWS_ACCOUNT_ID: "123456789012"
    AWS_REGION: us-east-1

agentbreeder deploy agent.yaml --target cloud-run --region us-central1

Set GOOGLE_APPLICATION_CREDENTIALS and GOOGLE_CLOUD_PROJECT in your environment.

agentbreeder deploy agent.yaml --target container-apps

agentbreeder deploy agent.yaml --target kubernetes

Works with EKS, GKE, AKS, and self-hosted clusters. Ensure your kubectl context is pointed at the target cluster.

agentbreeder deploy agent.yaml --target claude-managed

No container is built. Add ANTHROPIC_API_KEY to secrets:

deploy:
  cloud: claude-managed
  secrets:
    - ANTHROPIC_API_KEY

claude_managed:
  environment:
    networking: unrestricted
  tools:
    - type: agent_toolset_20260401

Framework-Specific Configuration

Some frameworks support extra config blocks in agent.yaml. These are optional — AgentBreeder picks sensible defaults.

Claude SDK

claude_sdk:
  thinking:
    type: adaptive          # "adaptive" (default) | "enabled"
    effort: high            # "low" | "medium" | "high"
  prompt_caching: true      # cache system prompt for latency + cost savings

Google ADK

google_adk:
  session_backend: memory           # "memory" | "database" | "vertex_ai"
  session_db_url: ""                # required if session_backend is "database"
  memory_service: memory            # "memory" | "vertex_ai_bank" | "vertex_ai_rag"
  artifact_service: memory          # "memory" | "gcs"
  gcs_bucket: ""                    # required if artifact_service is "gcs"

Claude Managed

claude_managed:
  environment:
    networking: unrestricted        # "unrestricted" | "restricted"
  tools:
    - type: agent_toolset_20260401  # full built-in toolset

Common Workflows

Push your agent to the registry

Before (or instead of) deploying, push the agent and its dependent entities to your org's registry so they can be re-used, versioned, and deployed from Studio.

# Push the prompt — versioned. Auto-bumps semver patch on update.
agentbreeder registry prompt push prompts/microlearning-system.md \
    --version 1.0.0 \
    --team eng \
    --description "Microlearning ebook author"

# Push the tools. The CLI auto-detects dispatch type:
#   .py file → python:<abs_path>
#   .ts/.js  → node:<abs_path>
#   module path → kept for in-process import
agentbreeder registry tool push tools/web_search.py
agentbreeder registry tool push tools/markdown_writer.py

# Push the agent record itself (refs the prompt + tools above).
# Works for both Python agents (top-level `framework:`) and polyglot
# agents (top-level `runtime: {language, framework}`) — the runtime
# block is validated at deploy time by the runtime registry, so the
# registry accepts any framework string under `runtime.framework`.
agentbreeder registry agent push agent.yaml

List anything in the registry:

agentbreeder registry prompt list --team eng
agentbreeder registry tool list
agentbreeder registry agent list --team eng

Auth

All registry commands are auth-gated. Set AGENTBREEDER_API_TOKEN to a JWT obtained from POST /api/v1/auth/login.

Test before deploy

Run a prompt or a tool against the registry directly from your terminal — no agent, no container build.

# Render a prompt with a real Gemini call (POST /api/v1/registry/prompts/{id}/render)
agentbreeder registry prompt try microlearning-system \
    --input "What is your job?" \
    --model gemini-2.5-flash

# Run a tool with structured args (POST /api/v1/registry/tools/{tool_id}/execute)
agentbreeder registry tool run web-search \
    --args '{"query": "agentbreeder docs"}'

agentbreeder registry tool run markdown-writer \
    --args '{"path": "out.md", "content": "# Hello"}'

Once the agent is deployed, you can chat with it through the same registry-aware CLI:

agentbreeder registry agent invoke microlearning-ebook-agent \
    --input "What is your job?" \
    --endpoint https://microlearning-ebook-agent-sizukgalta-uc.a.run.app \
    --token "$AGENT_MICROLEARNING_EBOOK_AGENT_TOKEN"

The CLI calls POST /api/v1/agents/{agent_id}/invoke, which proxies server-side and attaches the bearer token from --token or the AGENT_<UPPER_SNAKE>_TOKEN env var.

Check status and interact

agentbreeder status
agentbreeder logs my-agent --follow
agentbreeder chat my-agent

Search and inspect the registry

agentbreeder search "support"          # search across all agents, tools, prompts, models
agentbreeder list                      # list all your deployed agents
agentbreeder describe my-agent         # detailed info about one agent

Roll back

agentbreeder deploy agent.yaml --version 1.0.0

Schedule a recurring run

agentbreeder schedule my-agent --cron "0 9 * * 1-5"    # weekdays at 9am

Run evaluations

agentbreeder eval my-agent --suite evals/support.yaml

Manage secrets

agentbreeder secret set ZENDESK_API_KEY           # stores in your configured backend
agentbreeder secret list
agentbreeder secret delete ZENDESK_API_KEY

Supported secret backends: environment variables, AWS Secrets Manager, GCP Secret Manager, HashiCorp Vault.

Discover available models and MCP servers

agentbreeder scan                         # discovers local MCP servers + models
agentbreeder scan --provider ollama       # Ollama models
agentbreeder scan --provider openrouter   # OpenRouter models
agentbreeder scan --provider litellm      # LiteLLM gateway

Tear down

agentbreeder teardown my-agent

Eject to Full Code

When you need custom routing, state machines, or programmatic control:

agentbreeder eject my-agent --to code

This generates a Python or TypeScript SDK project from your agent.yaml. Your YAML is preserved as the source of truth — the SDK wraps it with full programmatic access. See Full Code →.

All Frameworks × All Deployment Targets

Every framework compiles to every deployment target. The full matrix — and the per-target governance caveats — live in the agent.yaml reference → All Frameworks × All Deployment Targets.

Next Steps

What	Where
All `agent.yaml` fields	agent.yaml Reference →
All CLI commands	CLI Reference →
MCP server setup	MCP Servers →
RAG & knowledge bases	RAG →
Versioned prompts	Prompts →
Multi-agent orchestration	Orchestration SDK →
Full Python/TypeScript SDK	Full Code →
No-code Studio UI	No Code →

Low Code — agent.yaml + CLI

On this page