GitHub - Agent-Field/CodeAF: Open-Source Software factory for Open Models
CodeAF is a Go-based coding harness from AgentField AI, built to run coding agents on open models cheaply and manage several projects from one window. It is in early preview and Apache 2.0 licensed.
- On the DeepSWE benchmark, using the same DeepSeek model, its senior-dev specialist solved more issues than nine other harnesses, including Claude Code and Codex, at the lowest cost per solved issue.
- It runs as a single small binary with nothing else to install, and one window shows every project, conversation, and running task on the machine.
- Work is split across a planner, worker, and checker model per task, and changes land on their own branch, never on main, only after they pass checks.
CodeAF is a coding harness from AgentField AI, written in Go as a single small binary under Apache 2.0, and its pitch is cost-efficient coding on open models. The headline claim is that it ranked first on DeepSWE across ten coding harnesses running the same model (DeepSeek V4 Flash), solving 113 real GitHub issues and posting the lowest cost per solved issue, ahead of Claude Code, Codex, OpenCode, Kilo, and DeepSeek's own harness. Those are the project's own benchmark numbers, so they're worth checking against the methodology before relying on them. The article is also candid that this is an early preview with rough edges.
The bigger idea is the workflow. Instead of one terminal per repository, CodeAF gives you a single window listing every project and conversation, with a "needs you" queue, unread markers, and running tasks across all of them. You hand off work in plain language, and tasks run in separate copies of the repo on their own branches, landing only when a check passes and never directly on main, dev, or a release branch. Unverifiable work waits for your accept or reject. Built-in specialists handle coding changes (senior-dev), local security audits (sec), and GitHub pull request review (review), with reports in Markdown, JSON, or SARIF. The review and sec specialists are imported from PR-AF and SEC-AF and only read code, not run it.
Model routing is handled by a "crew" of three seats per task (worker, planner, checker) picked from the models your providers serve, with per-task and daily spend limits, and an option to restrict everything to open-weight models. A part worth scrutinizing is the opt-out model pool: by default each install sends anonymized, text-free performance numbers about models to a shared index, which you can inspect with `codeaf pool status` or disable with `model_pool = off`. Scheduled and watched tasks are set up in natural language and run only while codeaf is open, catching up missed runs on startup. The practical question for a read-later decision is whether the cost-per-solved-issue advantage holds up on your own codebase, which is the one thing the article can't tell you.