AutoScientists introduces a novel approach to long-running scientific experimentation by deploying decentralized, self-organizing teams of AI agents. It addresses the limitations of single-trajectory or centrally planned agent systems by enabling agents to autonomously form teams around promising hypotheses, critically evaluate proposals before committing compute, and share learnings. This facilitates sustained parallel exploration and avoids redundant effort, significantly accelerating discovery across complex scientific domains. The system is designed for researchers and engineers aiming to automate and enhance computational scientific workflows.

How It Works

The core architecture features Claude Code subagents coordinated through a local ClawInstitute server. Agents self-organize into teams, fostering a collaborative critique process for experimental proposals prior to execution. This decentralized coordination, managed by a pure orchestrator that launches agents and collects results without direct training involvement, allows the system to sustain parallel search and adapt its exploration strategy as evidence accumulates over hours or days. This contrasts with traditional agent systems that follow a singular research path or rely on a central planner, enabling more robust and efficient long-term experimentation.

Quick Start & Requirements

• Prerequisites: Node.js 22+, Python 3.9+, Claude Code CLI (claude).
• Installation:
  • Start ClawInstitute server: npx clawinstitute start (or npm install -g clawinstitute for global install).
  • Install Python dependencies: pip install -r requirements.txt.
• Running: Execute tasks via the claude CLI, e.g., claude -p "Read runbook.md and execute. Task: task-autoresearch. Run name: ar_v1.". Each run creates a new sibling directory for system state and logs.
• Hardware: Varies per task; consult individual task-<name>/README.md files.
• Links: Task-specific READMEs contain details for setup and execution.

Highlighted Details

• BioML-Bench: Achieved 74.4% mean leaderboard percentile across 24 biomedical ML tasks, outperforming prior AI agents by 8.33%.
• nanoGPT Training: Reached target validation metrics 1.9x faster, with 7 accepted improvements compared to a single-agent baseline's zero.
• ProteinGym Fitness: Demonstrated +12.5% improvement on the ACE2-Spike binding assay and a +6.5% average across 217 assays.

Maintenance & Community

No specific details regarding contributors, sponsorships, or community channels (e.g., Discord, Slack) are provided in the README.

Licensing & Compatibility

The README does not specify a software license. Compatibility for commercial use or closed-source linking is therefore undetermined.

Limitations & Caveats

Hardware requirements are task-dependent and can be substantial. The orchestrator's role is strictly coordination, not direct training. The system is optimized for long-running, complex experiments, implying significant setup and execution time investment.