SWE-AF

About

SWE-AF exists to turn a single API call into a coordinated engineering organization rather than a single model session. Built on the AgentField control plane, it registers a swe-planner node that decomposes a product requirement into an architecture and an issue DAG, then runs hundreds of role-specialized agent instances — coders, QA, reviewers, merger, verifier — on isolated git worktrees with hardness-aware execution and explicit technical-debt tracking. Execution is organized as three nested adaptive control loops with per-role model assignment (a flat models map assigns Claude, MiniMax, or OpenRouter models per role), a post-PR CI gate loops on failing checks, and builds checkpoint so a crash resumes rather than restarts. A sub-harness endpoint (swe-planner.implement_issue) plus a bundled Claude Code skill let outer harnesses delegate single issues to it. The runtime is free Apache-2.0 software; users pay only the underlying model tokens, and typical full builds cost a few dollars to tens of dollars.