SWE-AF

SWE-AF is an open-source runtime from Agent-Field that turns one API call into a coordinated team of AI agents (product manager, architect, coder, reviewer, tester) which scope, build, and ship software. It is built on the AgentField control plane, is labeled public beta in its README, and is licensed under Apache 2.0.

What It Is

SWE-AF is an autonomous software engineering factory. You submit a goal plus a repository (a remote repo_url or a local repo_path), and it plans the work, generates and reviews an architecture, splits it into a dependency-sorted set of issues, and executes them in parallel in isolated git worktrees. The result is merged, integration-tested, and verified against acceptance criteria, optionally ending in a pull request.

It can be triggered with the af CLI (af call swe-planner.build) or via HTTP to the AgentField control plane. Multi-repository builds are supported, with a primary repo that drives the build and dependency repos whose failures do not block it.

How the Adaptive Control Loops Work

The README describes three nested loops. An inner loop retries a coder with QA or review feedback. A middle loop invokes an issue advisor that retries with a new approach, splits the work, or accepts with tracked debt. An outer loop uses a replanner to restructure the remaining issue DAG after escalated failures. Optional continual learning injects conventions and failure patterns found early into downstream issues, and checkpointed execution allows resume_build after interruptions.

Models, Runtimes, and Deployment

Runtimes include claude_code, open_code (OpenRouter, OpenAI, Google, Anthropic, and other gateway model IDs) and codex. Models can be assigned per role, such as a stronger coder and a lighter QA agent. Setup paths include af install from GitHub, a one-click Railway deploy, a local Python 3.12+ install, and Docker Compose with scalable workers. A Go implementation under go/ registers under the same node IDs.

Sub-Harness Mode

For callers that are already coding harnesses, implement_issue skips planning and runs only the coder and reviewer loop on an isolated issue/<build_id>-<slug> branch, returning the branch without touching the caller's checkout. After a PR is opened, a CI gate can watch GitHub Actions and run a bounded fix-and-repush loop. The README publishes its own benchmark and example-PR results, which are vendor-reported.