AgentField｜AI Agent 控制平面与服务网格

access：来源确认 · curl -fsSL https://agentfield.ai/install.sh | bash

project_type：来源确认 · AgentField is an open-source control plane that lets you build AI agents callable by any service in your stack

<div align="center">

# AgentField — The AI Backend

### **Build agents like APIs. Run ten thousand of them like microservices.**

*One request fans out to thousands of agents. The control plane queues, retries, and traces every branch.*

[![Stars](https://img.shields.io/github/stars/Agent-Field/agentfield?style=flat&logo=github&logoColor=e8e5dc&color=0c0b09&labelColor=8b7355)](https://github.com/Agent-Field/agentfield/stargazers)
[![License](https://img.shields.io/badge/license-Apache%202.0-0c0b09.svg?style=flat&labelColor=8b7355)](LICENSE)
[![Downloads](https://img.shields.io/endpoint?url=https%3A%2F%2Fgist.githubusercontent.com%2Fsantoshkumarradha%2Fd98e2ad73502b4075f6a5f0ae4f5cae5%2Fraw%2Fbadge.json&style=flat&logo=download&logoColor=e8e5dc&color=0c0b09&labelColor=8b7355&cacheSeconds=3600)](https://github.com/Agent-Field/agentfield)
[![Coverage](https://img.shields.io/endpoint?url=https%3A%2F%2Fgist.githubusercontent.com%2Fsantoshkumarradha%2F433fb09c2d54c3c2589125cfd3eb14a2%2Fraw%2Fbadge.json&style=flat&logo=github&logoColor=e8e5dc&color=0c0b09&labelColor=8b7355&cacheSeconds=3600)](https://github.com/Agent-Field/agentfield/actions/workflows/coverage.yml)
[![Last Commit](https://img.shields.io/github/last-commit/Agent-Field/agentfield?style=flat&logo=git&logoColor=e8e5dc&color=0c0b09&labelColor=8b7355)](https://github.com/Agent-Field/agentfield/commits/main)
[![Discord](https://img.shields.io/badge/discord-join%20us-0c0b09.svg?style=flat&labelColor=8b7355&logo=discord&logoColor=e8e5dc)](https://discord.gg/aBHaXMkpqh)

<!-- All outbound agentfield.ai links must use UTM parameters, not direct URLs. See assets/utm-links.csv for the full list — target URLs there are for verification and reference only, always use the UTM version in the README. Update both the README and the CSV when adding or changing links. -->
**[Docs](https://agentfield.ai/docs/learn?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-docs)** · **[Quick Start](https://agentfield.ai/docs/learn/quickstart?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-quickstart)** · **[Python SDK](https://agentfield.ai/docs/reference/sdks/python?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-python-sdk)** · **[Go SDK](https://agentfield.ai/docs/reference/sdks/go?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-go-sdk)** · **[TypeScript SDK](https://agentfield.ai/docs/reference/sdks/typescript?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-typescript-sdk)** · **[REST API](https://agentfield.ai/docs/reference/sdks/rest-api?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-rest-api)** · **[Examples](#built-with-agentfield)** · **[Discord](https://discord.gg/aBHaXMkpqh)**

</div>

<div align="center">
<a href="https://agentfield.ai/docs/build/intelligence/harness?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-harness-banner">
<img src="assets/harness-banner.png" alt="Now includes Harness Orchestration — multi-turn coding agents with CodeAF, Claude Code, Codex, Gemini CLI, and OpenCode" width="100%" />
</a>
</div>

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.

<div align="center">

https://github.com/user-attachments/assets/9fb7b1cf-26de-4b9b-9ba2-917252cc26ec

<sub><b>One prompt → a running containerized production ready multi-agent backend.</b> No glue code, start using the agent API!</sub>

</div>

## 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.

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

The installer also installs the `codeaf` coding harness in
`$AGENTFIELD_HOME/bin` (`~/.agentfield/bin` by default), so harness-backed agents
work out of the box; skip it with `--no-codeaf` (`--no-aforge` remains an alias). Staging uses
`~/.agentfield-staging`; an explicit `AGENTFIELD_HOME` selects the production
managed home. `AGENTFIELD_INSTALL_DIR` selects the `af` directory; CodeAF
always uses the managed home's `bin`. Repair it with `af codeaf ensure`.
The provider name `aforge` remains an alias; set `AFORGE_BIN` and select
`provider="aforge"` to pin the legacy adapter. CodeAF state stays isolated in
`$AGENTFIELD_HOME/codeaf`. See [harness providers](docs/harness-providers.md)
for binary overrides, mirror/skip switches, and telemetry controls.

On macOS the installer also registers the control plane to start at login (under
launchd) and adds a menu-bar icon. Stop it with `af service stop` or the menu-bar
icon — a plain `kill` looks like a crash and it restarts. `af service status`
shows health and in-flight work; install with `--no-tray` to skip this entirely.

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

```text
/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 →](https://agentfield.ai/docs/learn/build-with-claude-code?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-prompt-to-production)

## The DX you get

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

```python
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 →](https://agentfield.ai/docs/learn?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-read-full-docs)

<details>
<summary><b>Need approvals, audit trails, and governance? (the enterprise sample)</b></summary>

```python
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.
```

> **What you just saw:** `app.ai()` calls an LLM and returns structured output. `app.pause()` suspends for [human approval](https://agentfield.ai/docs/build/execution/human-in-the-loop?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-human-in-the-loop). `app.call()` routes to other agents through the control plane. `app.run()` auto-exposes everything as REST. [Read the full docs →](https://agentfield.ai/docs/learn?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-read-full-docs)

</details>

<details>
<summary><b>Prefer to scaffold by hand? (Python / Go / TypeScript / Docker)</b></summary>

```bash
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
```

```bash
# 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!"}}'
```

```bash
# 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
```

[Deployment guide →](https://agentfield.ai/docs/reference/deploy?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-deploy) for Docker Compose and production setups, or use the repository's [Kubernetes deployment guide](docs/deploying-on-kubernetes.md).

</details>

## See it in action

<!-- TODO: replace static UI.png with a short gif of a live fan-out run in the DAG view -->

<div align="center">
<img src="assets/UI.png" alt="AgentField Dashboard" width="100%" />
<br/>
<sub>Real-time workflow DAGs · Execution traces · Agent fleet management · Audit trails</sub>
</div>

## 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.

|  | **Frameworks**<br><sub>LangChain · CrewAI · PydanticAI · OpenAI Agents SDK</sub> | **Workflow engines**<br><sub>Temporal · Airflow</sub> | **Visual builders**<br><sub>n8n · Zapier</sub> | **AgentField** |
|---|:-:|:-:|:-:|:-:|
| Build agent logic (prompts, tools, structured output) | ● | — | — | ● |
| Prebuilt chains, retrievers, integrations | ● | — | ◐ | [◐](https://agentfield.ai/docs/integrations?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-integrations) |
| Production REST APIs out of the box | — | ◐ | ● | ● |
| Async + retries + webhooks | — | ● | ◐ | **●** |
| Memory scopes (global · actor · session · workflow) | ◐ | — | — | ● |
| Service discovery + cross-agent calls | — | — | — | **●** |
| Distributed agents (register from anywhere, one mesh) | — | ◐ | — | **●** |
| Coding agents as functions (Claude Code · Codex · CLI) | — | — | — | **●** |
| Agent identity, access policies, signed audit trails | — | — | — | ● |
| Fleet observability (DAGs · metrics · traces) | — | ◐ | — | ● |
| Multi-language SDKs (Python · Go · TypeScript) | ◐ | ● | — | ● |

<sub>● full · ◐ partial · — not the focus</sub>

**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 →](https://agentfield.ai/docs/learn/vs-frameworks?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-vs-frameworks)

## 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 is admitted into a bounded in-process queue with backpressure (`429`/`503` plus `Retry-After`). On graceful shutdown, in-flight executions are terminated with `status_reason` `control_plane_shutdown` rather than silently dropped.

| Property | What it means |
|---|---|
| Stateless Go control plane | Horizontal scaling behind a load balancer. Add replicas to add capacity. |
| Bounded in-process admission | Backpressure returns `429`/`503` with `Retry-After`; graceful shutdown records `control_plane_shutdown`. |
| Async execution | Webhooks and SSE, no timeout limits. A single run can go for hours or days. |
| Backpressure | Queue-depth limits and circuit breakers keep a fan-out from overwhelming downstream agents. |
| Routing overhead | Roughly 100-200ms per cross-agent hop. It matters when a branch does little work per hop, so keep hops coarse when latency is tight. |

Two examples already run at this load. The [deep-research engine](https://agentfield.ai/github/deepresearch/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-deepresearch-repo) fanned out 10,000+ agent invocations in one workflow. The [security auditor](https://agentfield.ai/github/sec-af/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-sec-af-repo) runs 250 coordinated agents per audit.

[Deployment guide →](https://agentfield.ai/docs/reference/deploy?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-deploy-scale) for Docker Compose, Kubernetes, and production setups.

## What You Get

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

- **[Reasoners & Skills](https://agentfield.ai/docs/build/building-blocks/reasoners?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-reasoners)** - `@app.reasoner()` for AI judgment, `@app.skill()` for deterministic code
- **[Structured AI](https://agentfield.ai/docs/reference/sdks/python?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-structured-ai)** - `app.ai(schema=MyModel)` → typed Pydantic/Zod output from any LLM
- **[Harness](https://agentfield.ai/docs/build/intelligence/harness?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-harness)** - `app.harness("Fix the bug")` dispatches multi-turn tasks to CodeAF, AgentField's own coding harness — no setup. Add `provider="claude-code"` (or `codex`, `gemini`, `opencode`, `pi`, `omp`) to orchestrate someone else's.
- **[Cross-Agent Calls](https://agentfield.ai/docs/build/coordination/cross-agent-calls?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-cross-agent-calls)** - `app.call("other-agent.func")` routes through the control plane with full tracing
- **[Discovery](https://agentfield.ai/docs/reference/sdks/python?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-discovery)** - `app.discover(tags=["ml*"])` finds agents and capabilities across the mesh. `tools="discover"` lets LLMs auto-invoke them.
- **[Memory](https://agentfield.ai/docs/build/coordination/shared-memory?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-memory)** - `app.memory.set()` / `.get()` / `.similarity_search()` - KV + vector search, four scopes, no Redis needed

**Scale** - Production infrastructure for non-deterministic AI.

- **[Async Execution](https://agentfield.ai/docs/build/execution/async?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-async-execution)** - Fire-and-forget with webhooks, SSE streaming, retries. No timeout limits - agents run for hours or days.
- **[Canary Deployments](https://agentfield.ai/docs/learn/features?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-canary-deployments)** - Traffic weight routing, A/B testing, blue-green deploys. Roll out agent versions at 5% → 50% → 100%.
- **[Human-in-the-Loop](https://agentfield.ai/docs/build/execution/human-in-the-loop?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-human-in-the-loop)** - `app.pause()` suspends execution for human approval. Crash-safe, durable, audited.
- **[Observability](https://agentfield.ai/docs/learn/features?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-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](https://agentfield.ai/docs/build/governance/identity?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-crypto-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](https://agentfield.ai/docs/build/governance/credentials?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-verifiable-credentials)** - Tamper-proof receipt for every execution. Offline-verifiable: `af vc verify audit.json`.
- **[Policy Enforcement](https://agentfield.ai/docs/build/governance/policy?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-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 →](https://agentfield.ai/docs/learn/features?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-full-features)

<div align="center">
<img src="assets/features-strip.png" alt="90+ Production Features" width="100%" />
</div>

<details>
<summary><h4 align="center">▼ Click to expand full capabilities</h4></summary>

#### AI & LLM

| Feature | How |
|---|---|
| Structured output (Pydantic/Zod) | `app.ai(schema=MyModel)` |
| Multi-turn coding agents | `app.harness("task")` — CodeAF by default |
| Orchestrate another harness | `app.harness("task", provider="claude-code")` (also `codex`, `gemini`, `opencode`, `pi`, `omp`) |
| LLM auto-discovers agents and tools | `app.ai(tools="discover")` |
| Multimodal (text, image, audio) | `app.ai("Describe", image_url="...")` |
| Streaming responses | `app.ai("...", stream=True)` |
| 100+ LLMs via LiteLLM | `AIConfig(model="anthropic/claude-sonnet-4-20250514")` |
| Temperature, max tokens, format | `app.ai(..., temperature=0.2)` |

#### Agent Mesh & Discovery

| Feature | How |
|---|---|
| Cross-agent calls with tracing | `app.call("agent.func", input={...})` |
| Discover agents by tag (wildcards) | `app.discover(tags=["ml*"])` |
| Discover by health status | `app.discover(health_status="active")` |
| Agent routers (namespacing) | `AgentRouter(prefix="billing")` |
| Auto context propagation | Workflow, session, actor IDs forwarded |
| Parallel agent execution | `asyncio.gather(app.call(...), ...)` |
| Auto-registration on startup | Service mesh with zero config |

#### Execution Engine

| Feature | How |
|---|---|
| Sync execution (REST) | [`POST /api/v1/execute/{agent}.{func}`](docs/api/EXECUTE.md) |
| Async (fire-and-forget) | [`POST /api/v1/execute/async/{agent}.{func}`](docs/api/EXECUTE.md) |
| Webhooks + HMAC-SHA256 signing | `AsyncConfig(webhook_url="...", secret="...")` |
| SSE streaming (real-time) | `/api/v1/execute/stream/{id}` |
| No timeout limits (hours/days) | Control plane allows unlimited duration |
| Execution polling | `GET /api/v1/executions/{id}` |
| Restart/replay | [`POST /api/v1/executions/{id}/restart`](docs/api/EXECUTION_RESTART.md) |
| Run naming, labels, links | [`POST /api/v1/runs/{id}/metadata`](docs/api/RUN_METADATA.md) |
| Batch status checks | `POST /api/v1/executions/batch-status` |
| Progress updates mid-execution | Intermediate payloads during long tasks |
| Auto retries + exponential backoff | Transparent - control plane handles |
| Backpressure + queue depth limits | Fair scheduling, circuit breakers |
| Bounded in-process queue | Backpressure and explicit graceful-shutdown termination |

#### Memory (Distributed State)

| Feature | How |
|---|---|
| Key-value storage | `app.memory.set(key, value)` / `.get(key)` |
| Vector search (semantic) | `app.memory.similarity_search(embedding, top_k=5)` |
| Four scopes | Global, actor, session, workflow |
| Reactive memory events | `@app.memory.on_change("order_*")` |
| Metadata filtering | Filter stored values by metadata |
| Zero dependencies | Built into control plane - no Redis |

#### Human-in-the-Loop

| Feature | How |
|---|---|
| Durable pause/resume | `await app.pause(reason="...")` |
| Approval workflows with UI | `approval_request_url` for reviewers |
| Configurable timeouts | `expires_in_hours=24` + auto-escalation |
| Crash-safe state | Survives agent restarts |

#### Canary Deployments & Versioning

| Feature | How |
|---|---|
| Traffic weight routing | 5% → 50% → 100% rollouts |
| A/B testing | 50/50 splits with `X-Routed-Version` |
| Blue-green deployments | Instant weight switch, zero downtime |
| Per-version health tracking | Unhealthy versions auto-removed |
| Agent lifecycle states | pending → starting → ready → degraded → offline |

#### Identity & Governance

| Feature | How |
|---|---|
| Cryptographic identity per agent | Auto-generated W3C DID + Ed25519 keys |
| Verifiable Credentials | Tamper-proof receipt per execution |
| Offline VC verification | `af vc verify audit.json` |
| Tag-based access policies | ALLOW/DENY rules on caller → target tags |
| Cryptographically signed requests | Ed25519 signatures on cross-agent calls |
| VC hierarchy (3 tiers) | Platform → Node → Function control |
| Agent notes (audit log) | `app.note("Decision", tags=["critical"])` |
| Non-repudiation | Cryptographic proof of actions |
| Permission request workflows | Auto-created when access denied |

#### Observability & Fleet Management

| Feature | How |
|---|---|
| Automatic DAG visualization | Workflow graphs in dashboard |
| Prometheus metrics | `/metrics` out of the box |
| Structured JSON logging | Automatic from SDK |
| Execution timeline | Chronological decision trace |
| Health checks (K8s-ready) | `/health`, `/ready` endpoints |
| Correlation IDs | `X-Workflow-ID`, `X-Execution-ID` |
| Workflow DAG API | `GET /api/v1/workflows/{id}/dag` |
| Agent heartbeat monitoring | Auto health status transitions |

#### Harness (Multi-turn Coding Agents)

| Feature | How |
|---|---|
| Zero-setup default harness | CodeAF (`codeaf`), installed alongside `af` |
| Swap the worker, keep the loop | `provider="claude-code"` \| `"codex"` \| `"gemini"` \| `"opencode"` \| `"pi"` \| `"omp"` |
| Fleet-wide default override | `AGENTFIELD_HARNESS_PROVIDER=codex` |
| Schema-constrained output | `schema=ResultModel` (Pydantic/Zod) |
| Cost capping | `max_budget_usd=3.0` |
| Turn limiting | `max_turns=100` |
| Tool access control | `tools=["Read", "Write", "Bash"]` |
| Environment injection | `env={"KEY": "value"}` |
| System prompt override | `system_prompt="..."` |
| OpenCode per-run configuration | All SDKs preserve caller `OPENCODE_CONFIG_CONTENT`; `AGENTFIELD_OPENCODE_STEPS` overrides the default `500` steps |
| OpenCode prompt compatibility | All SDKs support `AGENTFIELD_OPENCODE_INLINE_SYSTEM_PROMPT=1` for the opt-in inline rollback |
| Provider-agnostic reasoning variants | `variant="high"` or a `#high` model suffix |
| Multi-layer output recovery | Cosmetic repair → retry → full retry |

#### Connector API (Fleet Management)

| Feature | How |
|---|---|
| Remote agent management | `/connector/reasoners` |
| Version traffic control | `/connector/.../weight` |
| Bearer token auth | `AGENTFIELD_CONNECTOR_TOKEN` |
| Air-gapped deployment | Outbound WebSocket only |

#### Developer Experience

| Feature | How |
|---|---|
| CLI scaffolding | `af init my-agent --defaults --language python\|go\|typescript` |
| Local dev with dashboard | `af server` → http://localhost:8080 |
| Hot reload | `af dev` auto-detects changes |
| Auto-REST from decorators | Every `@app.reasoner()` → `POST /api/v1/execute/...` |
| Python, Go, TypeScript SDKs | Native patterns per language |
| MCP server integration | `af add --mcp --url <server>` |
| Zero-setup MCP for AI harnesses | Control plane serves MCP at `<server>/mcp` → [docs](docs/mcp-integration.md) |
| Config storage API | `POST /api/v1/configs/:key` - database-backed |
| Docker + Kubernetes ready | Stateless control plane, horizontal scaling |

[Explore all features in detail →](https://agentfield.ai/docs/learn/features?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-explore-features)

</details>

## Built With AgentField

Each of these is a real, installable agent node. With a control plane running, drop any of them into your setup with a single command — `af install <repo-url>` clones the repo, isolates its dependencies, and registers the node. You're prompted once for shared secrets like `OPENROUTER_API_KEY` (stored encrypted and reused across every node), then the node's reasoners are callable over REST or with `af call`:

```bash
af install https://github.com/Agent-Field/SWE-AF             # autonomous engineering team   → node: swe-planner
af install https://github.com/Agent-Field/sec-af             # security auditor              → node: sec-af
af install https://github.com/Agent-Field/cloudsecurity-af   # cloud / IaC security scanner  → node: cloudsecurity
af install https://github.com/Agent-Field/pr-af              # agentic code review           → node: pr-af
af install https://github.com/Agent-Field/person-to-brief-af # cited person intelligence     → node: person-to-brief-af

af run swe-planner                                          # start a node (prompts once for required secrets)
af call swe-planner.build --in '{"goal": "Add JWT auth", "repo_url": "https://github.com/user/my-repo"}'
```

Full walkthrough — authoring, installing, and configuring nodes: [Installing agent nodes →](docs/installing-agent-nodes.md).

<table>
  <tr>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/github/swe-af/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-swe-af-repo">
        <img src="assets/examples/autonomous-engineering-team.png" alt="Autonomous Engineering Team" />
      </a>
      <br/>
      <b>Autonomous Engineering Team</b>
      <br/>
      <sub>One API call spins up PM, architect, coders, QA, reviewers - hundreds of coordinated agents that plan, build, test, and ship.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/github/swe-af/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-swe-af-repo">View project →</a>
    </td>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/github/deepresearch/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-deepresearch-repo">
        <img src="assets/examples/deep-research-engine.png" alt="Deep Research Engine" />
      </a>
      <br/>
      <b>Deep Research Engine</b>
      <br/>
      <sub>Recursive research backend. Spawns parallel agents, evaluates quality, generates deeper agents, and recurses -10,000+ agents per query.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/github/deepresearch/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-deepresearch-repo">View project →</a>
    </td>
  </tr>
  <tr>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/github/mongodb/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-mongodb-repo">
        <img src="assets/examples/reactive-database-layer.png" alt="Reactive MongoDB Intelligence" />
      </a>
      <br/>
      <b>Reactive MongoDB Intelligence</b>
      <br/>
      <sub>Atlas Triggers + agent reasoning. Documents arrive raw and leave enriched - risk scores, pattern detection, evidence chains.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/github/mongodb/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-mongodb-repo">View project →</a>
    </td>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/github/sec-af/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-sec-af-repo">
        <img src="assets/examples/ai-security-auditor.png" alt="Autonomous Security Audit" />
      </a>
      <br/>
      <b>Autonomous Security Audit</b>
      <br/>
      <sub>250 coordinated agents trace every vulnerability source-to-sink and adversarially verify each finding. Confirmed exploits, not pattern flags.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/github/sec-af/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-sec-af-repo">View project →</a>
    </td>
  </tr>
  <tr>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/github/cloudsecurity/?utm_source=github-readme&utm_campaign=github-readme&utm_content=cloudsec&utm_id=github-readme-cloudsec-repo">
        <img src="assets/examples/cloud-security.png" alt="CloudSecurity AF" />
      </a>
      <br/>
      <b>CloudSecurity AF</b>
      <br/>
      <sub>AI-native cloud infrastructure security scanner that performs shift-left attack path analysis directly from IaC, prioritizing the most dangerous risk chains before deployment.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/github/cloudsecurity/?utm_source=github-readme&utm_campaign=github-readme&utm_content=cloudsec&utm_id=github-readme-cloudsec-repo">View project →</a>
    </td>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/github/pr-af/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-pr-af-repo">
        <img src="assets/examples/agentic-pr-reviewer-hero.png" alt="Agentic PR Reviewer — #1 open-source on Code-Review-Bench" />
      </a>
      <br/>
      <b>Agentic PR Reviewer</b>
      <br/>
      <sub>#1 open-source reviewer on Code-Review-Bench - 0.706 golden recall across 42 tools compared, at ~10x lower cost per review. Builds a custom review strategy for every PR, then adversarially challenges its own findings.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/github/pr-af/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-pr-af-repo">View project →</a>
    </td>
  </tr>
  <tr>
    <td align="center" width="50%">
      <a href="https://github.com/Agent-Field/person-to-brief-af">
        <img src="assets/examples/person-to-brief-af.png" alt="Person to Brief AF" />
      </a>
      <br/>
      <b>Person to Brief AF</b>
      <br/>
      <sub>Give it a name and company. It resolves the right identity, researches public professional sources in parallel, verifies claims, and returns an outreach-ready brief plus cited PDF.</sub>
      <br/><br/>
      <a href="https://github.com/Agent-Field/person-to-brief-af">View project →</a>
    </td>
    <td align="center" width="50%">
    </td>
  </tr>
</table>

[See all examples →](https://agentfield.ai/examples?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-see-all-examples)

Built something with AgentField? [Submit your project to be featured on the examples page](https://github.com/Agent-Field/agentfield/issues/new?template=community-project.md).

## Architecture

<div align="center">
<img src="assets/arch.png" alt="AgentField Architecture" width="100%" />
</div>

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 →](https://agentfield.ai/docs/learn/architecture?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-architecture)

## Learn More

AgentField sends privacy-minimized installation and lifecycle telemetry by default; no prompts or execution payloads are sent. Set `AGENTFIELD_TELEMETRY_ENABLED=false` to disable it. [Telemetry details](docs/ENVIRONMENT_VARIABLES.md#anonymous-telemetry).

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

<table>
  <tr>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/blog/what-is-harness-orchestration?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-what-is-harness-orchestration">
        <img src="assets/blog/what-is-harness-orchestration.webp" alt="What is harness orchestration?" />
      </a>
      <br/>
      <b>What is harness orchestration?</b>
      <br/>
      <sub>The atomic unit of intelligence is climbing from the model call to the autonomous harness - and what changes when it does.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/blog/what-is-harness-orchestration?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-what-is-harness-orchestration">Read post →</a>
    </td>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/blog/harness-as-black-box?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-harness-black-box">
        <img src="assets/blog/harness-black-box.webp" alt="Part 1: The Black Box" />
      </a>
      <br/>
      <b>Part 1: The Black Box</b>
      <br/>
      <sub>Treating harnesses like Claude Code and Codex as autonomous, embodied, persistent computational entities.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/blog/harness-as-black-box?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-harness-black-box">Read post →</a>
    </td>
  </tr>
  <tr>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/blog/harness-as-membrane?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-harness-membrane">
        <img src="assets/blog/harness-membrane.webp" alt="Part 2: Engineering the Membrane" />
      </a>
      <br/>
      <b>Part 2: Engineering the Membrane</b>
      <br/>
      <sub>Shaping the boundary surface of a harness across four engineerable dimensions: workspace, drift, verifier placement, and recovery budget.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/blog/harness-as-membrane?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-harness-membrane">Read post →</a>
    </td>
    <td align="center" width="50%">
      <a href="https://agentfield.ai/blog/ai-backend?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-ai-backend">
        <img src="assets/blog/ai-backend.webp" alt="The AI Backend" />
      </a>
      <br/>
      <b>The AI Backend</b>
      <br/>
      <sub>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.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/blog/ai-backend?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-ai-backend">Read post →</a>
    </td>
  </tr>
  <tr>
    <td align="center" width="50%">
      <b>Fan out 1,000 parallel agents from one request</b>
      <br/>
      <sub>A tutorial on turning a single call into a bounded fan-out across the control plane, with queues, retries, and traces on every branch.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/blog/fan-out-1000-agents?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-fan-out-1000-agents">Read post →</a>
    </td>
    <td align="center" width="50%">
      <b>Claude Code as a function</b>
      <br/>
      <sub>Wrap a multi-turn coding harness behind a REST endpoint and call Claude Code the same way you call any other agent.</sub>
      <br/><br/>
      <a href="https://agentfield.ai/blog/claude-code-as-a-function?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-blog-claude-code-as-a-function">Read post →</a>
    </td>
  </tr>
</table>

### Documentation

- **[vs Agent Frameworks](https://agentfield.ai/docs/learn/vs-frameworks?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-vs-frameworks)** - How AgentField compares to LangChain, CrewAI, and workflow engines
- **[Full Documentation](https://agentfield.ai/docs/learn?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-full-docs)**

## New from AgentField: CodeAF

<a href="https://agentfield.ai/github/codeaf/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-codeaf-repo">
  <img src="https://raw.githubusercontent.com/Agent-Field/CodeAF/dev/assets/readme/benchmark-deepswe.webp" alt="First on DeepSWE: senior-dev, CodeAF's developer subharness, solved the most tasks (54.9%) at the lowest cost per solved task (1x). Every other harness solved less and paid more per solve: mini-swe-agent 1.9x, codex 2.1x, pi 2.4x, claude-code 3.4x, omp, kilo and opencode about 4.5x, muse-code 11.3x, deepseek-harness 26.6x." width="100%">
</a>

**[CodeAF](https://agentfield.ai/github/codeaf/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-codeaf-repo)** is our new open-source coding harness,
built for open models: frontier-grade coding at a fraction of the cost. It ranked
**#1 on DeepSWE** out of ten coding harnesses on the same model (DeepSeek V4 Flash),
ahead of Claude Code, Codex and OpenCode, at the lowest cost per solved issue.
One Go binary, Apache 2.0.

[Try CodeAF →](https://agentfield.ai/github/codeaf/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-codeaf-repo) · [See the benchmark →](https://agentfield.ai/github/codeaf/?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-codeaf-benchmark#benchmarks)

## Community

<div align="center">

[![Discord](https://img.shields.io/badge/Join%20Discord-d4a24a?style=for-the-badge&logo=discord&logoColor=white)](https://discord.gg/aBHaXMkpqh)
[![Twitter](https://img.shields.io/badge/Follow%20on%20X-0c0b09?style=for-the-badge&logo=x&logoColor=white)](https://x.com/agentfield_ai)

**[GitHub Issues](https://github.com/Agent-Field/agentfield/issues)** · **[Documentation](https://agentfield.ai/docs/learn?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-community-docs)** · **[Examples](https://agentfield.ai/docs/learn/examples?utm_source=github-readme&utm_campaign=github-readme&utm_id=github-readme-community-examples)**

</div>

## License

[Apache 2.0](LICENSE)

需要 Python、Go 或 TypeScript 之一及对应 SDK；控制平面为 Go 服务，可 Docker 运行或通过安装脚本部署；调用 LLM 需自备模型服务与 API Key（README 示例使用 OPENROUTER_API_KEY 与 Anthropic 模型）；是否收费、企业功能授权、中国大陆网络可用性均待核验。