AgentField is an open-source control plane that lets you build AI agents callable by any service in your stack - frontends, backends, other agents, cron jobs - just like any other API. You write agent logic in Python, Go, or TypeScript. AgentField turns it into production infrastructure: routing, coordination, memory, async execution, and observability. Every function becomes a REST endpoint, and the same code scales from one agent on your laptop to ten thousand in a single workflow: the control plane handles the fan-out, the queues, and the retries.

Build production agents with a prompt.

Describe the system in one line. Get a production-ready multi-agent backend. Works in Claude Code, Codex, Gemini CLI, OpenCode, Aider, Windsurf, and Cursor.

curl -fsSL https://agentfield.ai/install.sh | bash

Then in your coding agent, paste any spec with /agentfield :

/agentfield Build a claims processor with risk scoring, pattern detection,
and human approval for low-confidence decisions.

You get a Docker Compose stack wired up end-to-end — the agent, the control plane, and a production ready REST API endpoint you can paste and curl into a terminal to try it. See it in action →

The DX you get

Plain Python (or Go / TypeScript) functions. No DSL, no YAML, no graph wiring.

import asyncio
from agentfield import Agent, AIConfig
from pydantic import BaseModel

app = Agent(
    node_id="researcher",
    version="1.0.0",# Canary deploys, A/B testing, blue-green rollouts
    ai_config=AIConfig(model="anthropic/claude-sonnet-4-20250514"),
)

class SubQuestions(BaseModel):
    questions: list[str]

@app.reasoner(tags=["research"])
async def research(question: str, depth: int = 0, model: str | None = None) -> dict:

    if depth >= 3:  # depth cap keeps fan-out bounded
        answer = await app.ai(system="Answer directly and concisely.", user=question, model=model)
        return {"question": question, "answer": answer}

    # Break the question into sub-questions
    plan = await app.ai(
        system="Break this into 3-5 independent sub-questions.",
        user=question, schema=SubQuestions, model=model,
    )

    # Fan out: each sub-question recurses on this same agent, through the control plane
    branches = await asyncio.gather(*[
        app.call(f"{app.node_id}.research", question=q, depth=depth + 1, model=model)
        for q in plan.questions
    ])

    # Synthesize the branches back into one answer
    synthesis = await app.ai(system="Synthesize these findings.", user=str(branches), model=model)
    return {"question": question, "answer": synthesis, "branches": branches}

app.run()
# This single line exposes: POST /api/v1/execute/researcher.research
# One request fans out to thousands of agents. The control plane queues, retries, and traces
# every branch. No broker, no queue setup, no timeout.

What you just saw: app.ai() calls an LLM and returns structured output. app.call() routes to other agents (or back to itself) through the control plane, so recursion becomes distributed fan-out. asyncio.gather() runs every branch in parallel. app.run() auto-exposes everything as REST. Read the full docs →

from agentfield import Agent, AIConfig
from pydantic import BaseModel

app = Agent(
    node_id="claims-processor",
    version="2.1.0",# Canary deploys, A/B testing, blue-green rollouts
    ai_config=AIConfig(model="anthropic/claude-sonnet-4-20250514"),
)

class Decision(BaseModel):
    action: str# "approve", "deny", "escalate"
    confidence: float
    reasoning: str

@app.reasoner(tags=["insurance", "critical"])
async def evaluate_claim(claim: dict) -> dict:

    # Structured AI judgment - returns typed Pydantic output
    decision = await app.ai(
        system="Insurance claims adjuster. Evaluate and decide.",
        user=f"Claim #{claim['id']}: {claim['description']}",
        schema=Decision,
    )

    if decision.confidence < 0.85:
        # Human approval - suspends execution, notifies via webhook, resumes when approved
        await app.pause(
            approval_request_id=f"claim-{claim['id']}",
            approval_request_url=f"https://internal.acme.com/approvals/claim-{claim['id']}",
            expires_in_hours=48,
        )

    # Route to the next agent - traced through the control plane
    await app.call("notifier.send_decision", input={
        "claim_id": claim["id"],
        "decision": decision.model_dump(),
    })

    return decision.model_dump()

app.run()
# This single line exposes: POST /api/v1/execute/claims-processor.evaluate_claim
# The agent auto-registers with the control plane, gets a cryptographic identity, and every
# execution produces a verifiable, tamper-proof audit trail.

af init my-agent --defaults                            # Scaffold agent
cd my-agent && pip install -r requirements.txt
af server          # Terminal 1 → Dashboard at http://localhost:8080
python main.py     # Terminal 2 → Agent auto-registers

# Call your agent
curl -X POST http://localhost:8080/api/v1/execute/my-agent.demo_echo \
  -H "Content-Type: application/json" \
  -d '{"input": {"message": "Hello!"}}'

# Go
af init my-agent --defaults --language go && cd my-agent && go run .

# TypeScript
af init my-agent --defaults --language typescript && cd my-agent && npm install && npm run dev

# Docker (control plane only)
docker run -p 8080:8080 agentfield/control-plane:latest

See it in action

_{Real-time workflow DAGs · Execution traces · Agent fleet management · Audit trails}

How AgentField fits in your stack

Most agent tools help you write agent logic. AgentField is what runs it in production: the layer that makes agents callable by other software, durable across failures, and observable when one request fans out to a thousand branches. Keep the framework you already use for authoring; a reasoner is a plain function, so existing LangGraph or CrewAI code can run inside one.

_{● full · ◐ partial · — not the focus}

Prototype in whatever you like. The moment a second service needs to call your agent, put it on AgentField. That is the point where you would otherwise start writing queues, retries, discovery, and tracing by hand.

Full comparison & decision guide →

How it scales

The control plane is a stateless Go service. You put more of them behind a load balancer and the fleet grows horizontally. Work lands in a durable PostgreSQL queue with lease-based processing, so a crash or a restart resumes where it left off instead of dropping the job.

Two examples already run at this load. The deep-research engine fanned out 10,000+ agent invocations in one workflow. The security auditor runs 250 coordinated agents per audit.

Deployment guide → for Docker Compose, Kubernetes, and production setups.

What You Get

Build - Python, Go, or TypeScript. Every function becomes a REST endpoint.

• Reasoners & Skills - @app.reasoner() for AI judgment, @app.skill() for deterministic code
• Structured AI - app.ai(schema=MyModel) → typed Pydantic/Zod output from any LLM
• Harness - app.harness("Fix the bug") dispatches multi-turn tasks to Claude Code, Codex, Gemini CLI, or OpenCode
• Cross-Agent Calls - app.call("other-agent.func") routes through the control plane with full tracing
• Discovery - app.discover(tags=["ml*"]) finds agents and capabilities across the mesh. tools="discover" lets LLMs auto-invoke them.
• Memory - app.memory.set() / .get() / .search() - KV + vector search, four scopes, no Redis needed

Scale - Production infrastructure for non-deterministic AI.

• Async Execution - Fire-and-forget with webhooks, SSE streaming, retries. No timeout limits - agents run for hours or days.
• Canary Deployments - Traffic weight routing, A/B testing, blue-green deploys. Roll out agent versions at 5% → 50% → 100%.
• Human-in-the-Loop - app.pause() suspends execution for human approval. Crash-safe, durable, audited.
• Observability - Automatic workflow DAGs, Prometheus /metrics, structured logs, execution timeline.

Govern - IAM for AI agents. Every agent gets a cryptographic identity. Identity, access control, and audit trails - built in.

• Cryptographic Identity - Every agent gets a W3C DID (decentralized identifier) - not a shared API key. Agents authenticate to each other the way services authenticate with mTLS, but with cryptographic signatures that travel with the agent.
• Verifiable Credentials - Tamper-proof receipt for every execution. Offline-verifiable: af vc verify audit.json.
• Policy Enforcement - Tag-based policy gates with cryptographic verification. "Only agents tagged 'finance' can call this" - enforced by infrastructure, not prompts.

See the full production-ready feature set →

Built With AgentField

Autonomous Engineering Team One API call spins up PM, architect, coders, QA, reviewers - hundreds of coordinated agents that plan, build, test, and ship. View project →	Deep Research Engine Recursive research backend. Spawns parallel agents, evaluates quality, generates deeper agents, and recurses -10,000+ agents per query. View project →
Reactive MongoDB Intelligence Atlas Triggers + agent reasoning. Documents arrive raw and leave enriched - risk scores, pattern detection, evidence chains. View project →	Autonomous Security Audit 250 coordinated agents trace every vulnerability source-to-sink and adversarially verify each finding. Confirmed exploits, not pattern flags. View project →
CloudSecurity AF AI-native cloud infrastructure security scanner that performs shift-left attack path analysis directly from IaC, prioritizing the most dangerous risk chains before deployment. View project →	Agentic PR Reviewer Builds a custom review strategy for every PR - spawns parallel reviewer agents with runtime-crafted prompts, adversarially challenges its own findings, and posts evidence-grounded inline comments. View project →

See all examples →

Built something with AgentField? Submit your project to be featured on the examples page.

Architecture

The control plane is a stateless Go service. Agents connect from anywhere - your laptop, Docker, Kubernetes. They register capabilities, the control plane routes calls between them, tracks execution as DAGs, and enforces policies. Full architecture docs →

Learn More

The thinking behind AgentField - essays on AI backends, harness orchestration, and the infrastructure production agents actually need.

What is harness orchestration? The atomic unit of intelligence is climbing from the model call to the autonomous harness - and what changes when it does. Read post →	Part 1: The Black Box Treating harnesses like Claude Code and Codex as autonomous, embodied, persistent computational entities. Read post →
Part 2: Engineering the Membrane Shaping the boundary surface of a harness across four engineerable dimensions: workspace, drift, verifier placement, and recovery budget. Read post →	The AI Backend Our thesis: in five years every serious software company will run an AI backend - a reasoning layer that makes the decisions that used to be hardcoded. Read post →
Fan out 1,000 parallel agents from one request A tutorial on turning a single call into a bounded fan-out across the control plane, with queues, retries, and traces on every branch. Read post →	Claude Code as a function Wrap a multi-turn coding harness behind a REST endpoint and call Claude Code the same way you call any other agent. Read post →

Documentation

• vs Agent Frameworks - How AgentField compares to LangChain, CrewAI, and workflow engines
• Full Documentation

Community

License

Apache 2.0