RISE-X-Lab/OpenCollab

OpenCollab – Multi‑Agent Coding Framework

What it is – OpenCollab is a Python‑based research framework that lets you define a team of coding agents (e.g., an analyst, a coder, a tester) and a workflow that coordinates how those agents hand‑off work to each other. The runtime enforces the declared collaboration topology, records every step, and measures whether the agents actually followed the plan (the paper calls this Adherence).

Why it matters – Modern code‑generation models (Claude Code, Codex, OpenHands, etc.) are usually run as a single monolithic agent. OpenCollab shows that a small, well‑engineered multi‑agent system (≈ 50 k lines of code) can outperform those larger code‑bases on standard software‑engineering benchmarks such as SWE‑bench Pro, Terminal‑Bench 2.1, and DeepSWE, while also giving researchers fine‑grained statistics about token usage, cost, and compliance with a prescribed collaboration graph.

Key features

• Team definition – YAML files describe each role’s prompt, model, tools, token budget, and who may message whom.
• Programmable workflows – Built‑in “Duo” (dual‑coder) and “Weave” (ordered session) workflows; users can author custom Python workflow modules.
• Controllable runtime – Enforces declared message edges, checks token budgets before every model call, and can output a JSONL trace of the whole run.
• Adherence metric – Records refusals and measures the percentage of runs that respect the declared topology (reported up to 97 %).
• CLI & SDK – uv run opencollab … for quick experiments; a Python API for deeper integration.
• Benchmark suite – Results page documents token‑cost trade‑offs and comparative scores against Claude Code, Codex CLI, and OpenHands.

Getting started

uv sync --locked                     # install locked dependencies
cp configs/.env.example configs/.env # add your OpenAI/Anthropic key
uv run opencollab --workspace .      # launch the default self‑collaboration team

You can then edit the supplied team.yaml or write a new workflow, e.g.:

uv run opencollab workflow run duo \
  --workspace /my/repo \
  --task 'Fix the public issue described here.' \
  --args '{"allow_unisolated_shell":true}'

Documentation & resources – The repo ships a full package guide, configuration guide, workflow tutorials (Duo, Weave), benchmark results, and a separate evaluation repo (OpenCollab‑Eval) for running on larger software‑engineering datasets.

Status – The latest pre‑release is v0.9.3 (Oct 2026). The authors plan a 1.0.0 release after their arXiv paper is accepted. The project has already won the OAIC Agentic Software Factory Hackathon and is supported by a youth open‑source fund.

License – Mulan Permissive Software License v2 (MulanPSL‑2.0).