AutoResearch — Autonomous AI Research Loop

Orchestrates end-to-end AI research using a two-loop architecture. The outer loop steers direction; the inner loop executes experiments rapidly. Outputs include structured findings, analysis notebooks, and draft papers.

Architecture

Outer Loop (strategic)
  └─ sets hypothesis, reads inner-loop summaries, adjusts direction
     Inner Loop (tactical)
       └─ runs experiment → collects metrics → writes summary → repeats

Phase 0 — Frame the Research Question

1. Clarify the research objective with the user (one clear hypothesis).
2. Identify baselines, datasets, and evaluation metrics.
3. Create research/plan.md with: hypothesis, success criteria, compute budget.
4. Search related work: use WebSearch for prior art.

Phase 1 — Inner Loop: Rapid Experimentation

For each experiment iteration:

1.1 Design the experiment

• Pick one variable to change (ablation-friendly).
• Document the change in research/experiments/<run_id>/config.md.

1.2 Execute

# Run experiment and capture output
python experiments/run.py --config research/experiments/<run_id>/config.md \
  2>&1 | tee research/experiments/<run_id>/log.txt

1.3 Record metrics

Write results to research/experiments/<run_id>/metrics.json:

{
  "run_id": "<run_id>",
  "hypothesis": "...",
  "metric_name": 0.0,
  "wall_time_s": 0
}

1.4 Write summary

Append one-paragraph summary to research/findings.md:

• What changed, what the metric showed, what it implies.

1.5 Decide: continue or surface to outer loop

• Continue if the trend is clear and budget remains.
• Surface if results are surprising, budget is 80% spent, or a pattern emerges.

Phase 2 — Outer Loop: Synthesis and Steering

Triggered after each inner-loop surface event.

1. Read all research/experiments/*/metrics.json and research/findings.md.
2. Identify the top-performing configuration and the clearest pattern.
3. Generate a ranked hypothesis list for the next inner-loop batch.
4. Update research/plan.md with revised direction.
5. If the research objective is met → proceed to Phase 3.

Phase 3 — Output Generation

Research Report

Create research/report.md containing:

• Abstract (3–5 sentences)
• Introduction with motivation
• Methods: what was varied, how experiments were run
• Results: tables of key metrics, winning configuration
• Analysis: what patterns emerged, why
• Conclusion and future work

Presentation (optional)

If the user wants slides, generate them with a presentation tool using research/report.md as input.

Guardrails

• Never delete experiment logs — always append.
• Always validate that code runs before marking an experiment complete.
• Cap inner-loop iterations at the compute budget in research/plan.md.
• Reproducibility: log random seeds, library versions, hardware specs in each config.

Output Files