#201201st of 236 in Coding agents and AI editors

swe-af

Multi-agent runtime that plans, codes, reviews and merges repo changes

Stars
1.1k
License
Apache-2.0
Last commit
Nov 2026
Language
Go
Stars, last 29 weeks+429
25 Apr 2026: 703 stars2 May 2026: 727 stars9 May 2026: 737 stars16 May 2026: 762 stars23 May 2026: 782 stars30 May 2026: 806 stars6 Jun 2026: 826 stars13 Jun 2026: 839 stars20 Jun 2026: 868 stars27 Jun 2026: 871 stars4 Jul 2026: 884 stars11 Jul 2026: 906 stars18 Jul 2026: 919 stars25 Jul 2026: 941 stars1 Aug 2026: 950 stars8 Aug 2026: 961 stars15 Aug 2026: 970 stars22 Aug 2026: 978 stars29 Aug 2026: 984 stars5 Sept 2026: 988 stars12 Sept 2026: 990 stars19 Sept 2026: 1,005 stars26 Sept 2026: 1,009 stars3 Oct 2026: 1,022 stars10 Oct 2026: 1,031 stars17 Oct 2026: 1,033 stars24 Oct 2026: 1,041 stars31 Oct 2026: 1,122 stars7 Nov 2026: 1,125 stars14 Nov 2026: 1,132 stars

703 on 25 Apr 20261,132 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

SWE-AF runs on the AgentField control plane and registers a `swe-planner` node. A `build` call with a goal and a repo URL or path starts planning, coding, review, QA and merge agents, each coder in its own git worktree. Failed issues go through retry, issue-advisor and DAG-replanning loops. Runtimes include AForge, Claude Code, OpenCode, Codex and Command Code, and one build can change several repos.

Who it is for: Teams running autonomous multi-agent coding builds on their own AgentField setup

Strengths

  • Three nested control loops: coder retry, issue advisor, and replanning of the remaining DAG
  • Per-role model mapping (coder, qa, architect) across Claude, OpenRouter, OpenAI and Google
  • Parallel coders in isolated git worktrees; checkpointed builds can resume with resume_build
  • Multi-repo builds with primary and dependency roles in a single call

Weaknesses

  • Requires an AgentField control plane plus an LLM provider key; not standalone
  • Labeled public beta and has no tagged release
  • Cost scales with agent count: the example PR used 79 invocations and $19.23
  • Command Code runtime is not supported on the Go node

What it needs

  • no GPU
  • Docker + Compose
  • Needs AgentField control plane (af), Python 3.12+, LLM provider API key (Anthropic, OpenRouter, OpenAI or Google), GH_TOKEN for private repos and PRs, PostgreSQL (Railway deploy)
  • Models: Claude, OpenRouter models, OpenAI Codex, Google, MiniMax
  • port 8080

Also in Coding agents and AI editors

See all 236
Also in Coding agents and AI editors
RankProjectScore
1claw-codeRust CLI coding agent harness using Anthropic and OpenAI-compatible APIs194.8k starsMIT96 out of 100
2t3codeDesktop, web and mobile front end for controlling local coding agent CLIs Live demo ↗ (opens in a new tab)64.4k stars+36,149 (+128%) in 30 daysMIT96 out of 100
3codewhaleTerminal coding agent in Rust that works with 40+ model providers41.2k starsMIT94 out of 100
#4gooseRust AI agent with desktop app, CLI and API, extended through MCP56.7k stars+1,476 (+2.7%) in 30 daysApache-2.093 out of 100
#5ArchonYAML workflow engine that runs AI coding agents through fixed dev steps23.9k starsMIT93 out of 100