Arbor is an autonomous research agent designed to tackle long-horizon objectives by iteratively refining hypotheses. It assists researchers and engineers by automating code editing, experiment execution, and learning from results, enabling cumulative progress rather than isolated attempts. Its core benefit is a structured, persistent exploration framework that leverages past insights for smarter future research directions.

How It Works

Arbor employs a hypothesis-tree refinement approach, where each idea branches off, pruned if unsuccessful or harvested if valuable. It utilizes two agents: a Coordinator to manage the "Idea Tree" and drive the search, and an Executor to implement and run experiments in isolated Git worktrees. The process follows a six-step "arbor cycle": Observe, Ideate, Select, Dispatch, Backpropagate, and Decide. Experiments rigorously use separate development and held-out test splits, merging only gains that pass a configurable validation threshold to prevent overfitting.

Quick Start & Requirements

• Installation: Clone the repository, set up a Python virtual environment (≥ 3.10 recommended), and install via pip install -e .. The arbor doctor command verifies the setup.
• Prerequisites: Python version 3.10 or higher, and Git.
• Initial Setup: Run arbor setup to configure LLM providers and API keys, then arbor to start an interactive research session.
• Documentation: Links to the project page, paper, and documentation are available.

Highlighted Details

• General-purpose optimization: Capable of optimizing diverse tasks including model training, harness engineering, and data synthesis.
• Flexible runtime: Offers a robust native CLI runtime (recommended) and an Agent Skill Suite for integration into environments like Codex or Claude Code.
• Structured exploration: The Idea Tree framework ensures insights persist and propagate, guiding future research.
• Rigorous evaluation: Employs isolated Git worktrees and dedicated branches for experiments, with strict dev/test split validation to ensure robust improvements.
• Broad LLM support: Integrates with Anthropic, OpenAI, and various OpenAI-compatible backends via LiteLLM.
• Steerable interface: Features a live terminal dashboard, read-only WebUI, optional human-in-the-loop review, and domain-specific plugins.
• Performance: Demonstrates significant gains across multiple benchmarks, outperforming strong single-agent baselines.

Maintenance & Community

Information regarding specific maintainers, community channels (e.g., Discord, Slack), or roadmaps is not detailed in the provided README.

Licensing & Compatibility

• License: Released under the Apache License 2.0.
• Compatibility: The Apache 2.0 license is permissive, generally allowing for commercial use and integration into closed-source projects.

Limitations & Caveats

The provided README does not explicitly detail limitations, unsupported platforms, or known bugs.