SWE AF

SWE AF is a self-hosted system that assigns software development work to a team of AI agents. Built on AgentField, it suits developers who want to delegate changes across a repository or several connected codebases, with agents handling planning, coding, review and testing through to a pull request.

The orchestration runs locally or on your server, including through Docker; Railway is another deployment option. Model requests go to external AI providers and require provider credentials and an internet connection. The project is open source under Apache 2.0.

Its main distinction is how it responds when work gets stuck. A failed review feeds corrections back to the coder. Persistent failures can lead to a different approach, a smaller task, or a revised plan for the remaining work. If agents accept reduced scope, the system records the resulting technical debt and its severity.

Agents can work in parallel in separate Git worktrees, while scheduling respects dependencies between tasks. This supports changes that span an application and its shared libraries or dependent services. Reviewers and testers check individual issues, and a final verification checks the result against acceptance criteria.

SWE AF supports Claude Code, OpenCode and the OpenAI Codex CLI. You can assign different models to coding, architecture and QA roles, using providers such as Anthropic, OpenRouter, OpenAI and Google. Optional learning carries repository conventions and earlier failure patterns into later tasks. Checkpoints let interrupted builds resume.