AgentField AI Backend

What does this agent do, and when should you use it?

AgentField is an open-source control plane for running AI-agent logic written in Python, Go, or TypeScript as services. Developers define functions with `@app.reasoner()` or `@app.skill()`, and `app.run()` exposes them as REST endpoints while registering the node with the control plane. The control plane is a stateless Go service that routes calls between agents, tracks workflow DAGs, processes queues, retries work, and enforces policies; its documentation describes a PostgreSQL-backed lease-based durable queue. The project also documents memory, asynchronous webhooks and SSE, human approval, observability, and weighted version deployment. It can be deployed locally, with Docker, or on Kubernetes, making it aimed at teams operating coordinated agents as backend services.

An application creates a node with Agent(...), then registers AI reasoning functions with @app.reasoner() or deterministic code with @app.skill(). Those functions can use app.ai() for LLM calls and structured Pydantic/Zod output, app.call("other-agent.func") to route work through the control plane, and app.discover(tags=[...]) to find node capabilities. app.run() automatically serves endpoints such as POST /api/v1/execute/{agent}.{func}; asynchronous work can use POST /api/v1/execute/async/{agent}.{func} and report through webhooks, SSE, or execution-status endpoints. Long-running work can pause with app.pause() for human approval, while the control plane records execution DAGs, structured logs, metrics, and audit-related data. The CLI can install and register existing nodes with af install <repo-url>, start them with af run, and invoke them with af call.

1. A backend team wants a Python research function to break questions into subquestions and distribute them concurrently across nodes with app.call().
2. An insurance or risk team needs a low-confidence decision to suspend with app.pause() until a human reviewer approves it.
3. A platform team wants agents running on laptops, in containers, or on Kubernetes to auto-register and communicate through REST APIs.
4. A developer-tools team wants to expose multi-turn coding work from Claude Code, Codex, Gemini CLI, or OpenCode through app.harness().
5. An operations team needs long-running asynchronous tasks with webhooks, SSE, execution DAGs, and Prometheus /metrics for visibility.

What are this agent's strengths and limitations?

How do you install or deploy this agent?

For the documented Python path:

af init my-agent --defaults
cd my-agent && pip install -r requirements.txt

In terminal 1: af server
In terminal 2: python main.py
The CLI can also be installed with: curl -fsSL https://agentfield.ai/install.sh | bash. After initializing a Go node, run go run .; after initializing a TypeScript node, run npm install && npm run dev. The basic example does not specify required credentials. When installing a node with af install, the CLI prompts for required shared secrets, such as OPENROUTER_API_KEY.

How do you use this agent?

Once the control plane and node are running, invoke the documented example endpoint:

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

An existing node can be installed, started, and called as follows:

af install https://github.com/Agent-Field/SWE-AF
af run swe-planner
af call swe-planner.build --in '{"goal": "Add JWT auth", "repo_url": "https://github.com/user/my-repo"}'

To author a node, define an Agent, add @app.reasoner() functions, and call app.run().

How does this agent compare with similar options?

FAQ