Agent-Field/CodeAF

CodeAF – an open‑source “software factory” for coding with LLMs

What it is – CodeAF is a Go‑based command‑line tool that lets you run LLM‑powered coding agents across all the repositories on a machine. It presents a single “home” window that lists every project, its active conversations, pending tasks and spend. From there you can open a project in a tab, hand off a natural‑language request, and the system automatically:

Plans the work, splits it into independent tasks, and creates a branch for each.
Runs the work on a crew of three model seats (planner, worker, checker) chosen from a pool of open‑weight models (DeepSeek, Qwen, GLM, Kimi, MiniMax, Ollama, etc.).
Checks the results against configurable tests and merges only the passing changes.
Prompts you only when a decision is required (e.g., “which spend ledger should we use?”).

The result is a “factory” that can juggle many bugs, features or refactors at once while you intervene only where judgment is needed.

Key concepts

Concept	What it does
Home / tabs	One terminal window shows every repo, its tasks and cost. You can jump between projects without opening new terminals.
Conversation → tasks	Write a natural‑language request; CodeAF parses it into separate tasks, each on its own branch.
Sub‑harnesses	Specialist agents for recurring jobs (e.g., PR review, dependency triage). They have their own model plan, checks and typed I/O.
Crew / seats	For each task three model roles are selected automatically: planner (designs the plan), worker (writes code), checker (runs tests). You can pin a model to a seat.
Model pool	The tool measures latency, cost and success of each model on your install and uses that data to pick the cheapest/fastest model for a given seat.
Headless mode	`codeaf do …` runs a one‑off request from scripts/CI, returning JSON and an exit code.
Remote / anywhere	The UI runs locally while the conversation lives on the machine that owns the code. You can attach via SSH from a laptop, desktop or phone.
Telemetry	Anonymous usage counts (version, OS, token bands, errors) are sent by default; can be disabled via settings or env vars.

Installation & usage

curl -fsSL https://agentfield.ai/get/codeaf | bash   # installs a single binary (~53 MB)
codeaf                                            # start the UI

Pin a version with VERSION=<tag> or build from source (make build).

Typical workflow:

codeaf home – see all projects.
enter <project> – open a tab for a repo.
Type a request, e.g. fix the flaky test that fails 1/5 runs.
The UI creates a task tree; you approve/reject individual subtasks.
Accepted changes land on a new branch; you merge when ready.

Benchmarks & performance

DeepSWE benchmark – the built‑in /senior-dev sub‑harness solved the highest percentage of 113 real GitHub issues (54.9 %) on DeepSeek V4 Flash, at the lowest cost per solved issue, beating Claude Code, Codex, Kilo, etc.
Resource footprint – single binary 53 MB, ~27 MB RAM per session, 16 idle sessions use ~0.5 GB. This is 4‑21× smaller than competing agents.

Full benchmark details live under docs/benchmarks/.

When you might use CodeAF

Multi‑repo development – teams that need to juggle bugs, features and CI across many services.
Cost‑conscious LLM usage – automatically picks the cheapest model that meets quality requirements.
Automation – headless codeaf do … can be called from CI pipelines to apply bulk fixes (e.g., dependency upgrades).
Extensibility – you can author your own sub‑harnesses for domain‑specific tasks.

Limitations (as of the early preview)

Still in rapid development – expect rough edges and occasional crashes.
Requires an OpenRouter or provider API key for most models; Ollama works locally.
Telemetry is on by default (can be disabled).
The UI is terminal‑based; no graphical IDE integration yet.

Bottom line – CodeAF is a genuine open‑source project that turns LLMs into a coordinated, multi‑project coding assistant, emphasizing cost efficiency, task automation and a single‑window workflow. It is aimed at developers who want to harness open‑weight models without juggling many terminals or manual prompt engineering.