Arbor

Arbor is an open-source autonomous research system from the Gaoling School of Artificial Intelligence at Renmin University of China and Microsoft Research. Given a benchmark and a goal, it proposes hypotheses, edits code, runs experiments, and keeps only improvements that survive held-out evaluation. It ships as a Python CLI (arbor) and as an Agent Skill Suite for Codex and Claude Code.

What It Is

Arbor targets what its authors call Autonomous Optimization: an agent receives an initial artifact, an objective, a development evaluator and a held-out evaluator, then improves the artifact through iterative experimentation. Typical targets named in the project include model training, harness engineering, and data synthesis.

Two cooperating agents do the work. A long-lived Coordinator maintains the hypothesis (Idea) tree, and short-lived Executors implement single ideas in isolated git worktrees and report evidence back. Verified improvements merge into a per-run trunk, and the user promotes the trunk into main.

How the Arbor Cycle Works

Each cycle has six steps: Observe, Ideate, Select, Dispatch, Backpropagate, and Decide. Results and distilled insights are written to tree nodes and propagated to ancestors, so later proposals are conditioned on what was validated, pruned, or shown to have boundary conditions. Executors iterate on a dev split, but a change merges only if it clears a configurable margin on the held-out test split.

The project's paper and README report held-out wins over Codex and Claude Code on six tasks and an 86.36% Any-Medal rate on MLE-Bench Lite with GPT-5.5. These are the authors' own reported results.

Setup and Ways to Use It

Install with pip install arbor-agent (Python 3.10 or later and Git). Run arbor replay --demo to watch a recorded run without an API key, arbor quickstart or arbor setup to configure a provider, then arbor to start an intake conversation that produces a Research Contract. A keyless mode (arbor install plus arbor mcp) lets Claude Code or Codex use their own model while Arbor supplies deterministic tools.

Models, Search, and Steering

Arbor supports Anthropic, OpenAI Responses, and OpenAI-compatible backends through LiteLLM. An optional alphaXiv-backed literature search and arbor idea-check command assess an idea's novelty before experiments run. A live dashboard, read-only web UI, optional human-in-the-loop review modes, domain plugins, and resumable sessions with REPORT.md output let users steer and audit runs.