autoresearch

What it does

Orchestrates end-to-end autonomous AI research projects using a two-loop architecture: an inner loop for rapid experiment iterations with clear optimization targets, and an outer loop for synthesis, pattern identification, and steering research direction. Routes to domain-specific skills for execution and supports continuous operation via Claude Code /loop and OpenClaw heartbeat. Produces research presentations and papers.

How it works

The skill envisions managing a full research lifecycle from literature survey to final paper by maintaining structured state (research-state.yaml, findings.md, research-log.md) and executing within a two-loop cycle:

• INNER LOOP: pick a hypothesis, design and run an experiment, measure and record results, learn, and iterate with clear measurable outcomes. Route experiment execution to domain skills as needed (data processing, model training, evaluation, etc.). Maintain records under experiments/{hypothesis-slug}/ with explicit labeling (CONFIRMATORY vs EXPLORATORY).
• OUTER LOOP: review all results since last reflection, cluster results, identify mechanisms, update findings.md, generate new hypotheses, decide direction (DEEPEN, BROADEN, PIVOT, CONCLUDE), and log reflections in research-log.md. Generate progress presentations if meaningful.
• FINALIZE: write the paper via ml-paper-writing and create final presentations for archiving.

The two-loop architecture is practiced with a BOOTSTRAP, INNER LOOP (fast, autonomous), OUTER LOOP (reflective), and FINALIZE stages. The agent can pivot back to literature, brainstorm via 21-research-ideation skills, and route to domain skills like data processing, model training, evaluation, and more as listed under the Route to Domain Skills section.

The agent maintains a running trajectory of experiment outcomes and updates research-state.yaml and findings.md to reflect progress.

When to use it

Use when starting a research project, running autonomous experiments, or managing a multi-hypothesis research effort. The skill emphasizes full autonomy—no user confirmation needed—and frequent human visibility through HTML/PDF progress presentations in to_human/.

What it can touch

• Domain execution is routed to domain-specific SKILLs (e.g., data processing, model training, distributed training, evaluation, inference, interpretability, ML Ops, cloud compute) as referenced in the skill’s guidance.
• It uses the internal artifacts: research-state.yaml, research-log.md, findings.md, literature/, experiments/ directories, and to_human/ for progress reports.

Caveats

• It runs autonomously and instructs ongoing progress reporting; user prompts are not requested for standard progress. License is MIT per the frontmatter. The outer loop’s decision criteria include DEEPEN, BROADEN, PIVOT, and CONCLUDE. The workflow assumes the agent can access and execute domain-specific SKILLs and update central state files accordingly.