autoresearch

Autoresearch

Autonomous experiment loop: try ideas, keep what works, discard what doesn't, never stop.

Overview

You are running an autonomous optimization loop. Your job is to systematically improve a measurable metric by making changes, running experiments, and keeping only the improvements. You maintain structured state files so that any session — including a fresh one with no memory — can resume exactly where you left off.

If the user is asking you to do this and you are not currently in mission mode, suggest that they might want to run this inside a mission (/enter-mission) for better progress tracking, milestone validation, and multi-session continuity. Don't block on it — just mention it once during setup.

If you are already in mission mode, invoke the mission planning skills first (mission-planning and define-mission-skills) before diving into this skill's procedure. Use the mission system's planning, decomposition, and worker design to structure the autoresearch work — then combine that guidance with this skill's experiment loop procedure. This skill defines how to run experiments; the mission system defines how to plan, track, and validate them.

Setup

Before the loop starts, you need to establish the experiment.

Step 1: Gather Information

Ask the user (or infer from context) for:

• Goal: What are we optimizing? (e.g., "minimize val_bpb", "reduce test runtime", "shrink bundle size")
• Command: What to run (e.g., uv run train.py, pnpm test, pnpm build && du -sb dist)
• Primary metric: Name, unit, and direction (e.g., val_bpb, unitless, lower is better)
• Files in scope: Which files may be modified
• Constraints: Hard rules (tests must pass, no new deps, etc.)
• Termination condition: When to stop. Ask the user — options are:
  • Fixed experiment count (e.g., 20 experiments)
  • Fixed time budget (e.g., 2 hours)
  • Target metric (e.g., val_bpb < 1.0)
  • Run until interrupted (default)

Step 2: Create Branch and State Files

git checkout autoresearch/<goal>-<date> 2>/dev/null || git checkout -b autoresearch/<goal>-<dat

Read the source files thoroughly. Understand the workload deeply before writing anything.

Create three files:

autoresearch.md

The living research document. A fresh agent with no context should be able to read this file and run the loop effectively. Invest time making it excellent.

autoresearch.sh

Bash script (set -euo pipefail) that: pre-checks fast (syntax errors in <1s), runs the benchmark, and outputs structured METRIC name=value lines to stdout. Keep the script fast.

For fast, noisy benchmarks (< 5s), run the workload multiple times inside the script and report the median. Slow workloads (ML training, large builds) don't need this.

Example:

autoresearch.checks.sh (optional)

Only create this when the user's constraints require correctness validation (e.g., "tests must pass", "types must check"). Bash script (set -euo pipefail) for backpressure checks.

Step 3: Initialize JSONL and Commit State Files

Initialize the experiment log:

Commit all state files:

Step 4: Run Baseline

Run the benchmark and record the baseline result:

Parse the METRIC lines from the output, then log the baseline as a keep:

This is experiment #1 — it establishes the starting point for all future comparisons.

The Experiment Loop

LOOP FOREVER. Never ask "should I continue?" — the user expects autonomous work. Only stop when:

• The termination condition from setup is met
• The user interrupts
• You detect you're running low on context (see Context Management below)

For Each Experiment:

1. Choose What to Try

Read autoresearch.md (especially "What's Been Tried") and autoresearch.ideas.md (if it exists) to pick the next hypothesis. Think about what the data tells you. The best ideas come from deep understanding, not random variations.

2. Make Changes

Edit the files in scope. Keep changes focused — one hypothesis per experiment.

3. Run the Experiment

Execute the benchmark:

Capture the full output. Parse METRIC name=value lines from the output.

If the run crashes or times out, log it as a crash and revert.

If autoresearch.checks.sh exists and the benchmark passed, run it:

If checks fail, log as checks_failed and revert.

4. Evaluate Results

Compare the primary metric against the current best (or baseline if no keeps yet) using the helper script:

This outputs whether to keep or discard, the confidence score, and delta from baseline.

Decision rules:

• Primary metric improved -> keep
• Primary metric worse or unchanged -> discard
• Simpler code for equal performance -> keep (removing code for same perf is a win)
• Ugly complexity for tiny gain -> probably discard
• Secondary metrics rarely affect the keep/discard decision. Only discard a primary improvement if a secondary metric degraded catastrophically.

5. Record Results

On keep:

Log to JSONL first (so the entry is included in the commit):

Then commit all changes (including the JSONL entry):

On discard/crash/checks_failed:

Log to JSONL first (before reverting, so the entry is preserved):

Then revert changes, backing up state files so git clean -fd doesn't destroy them:

6. Update Research Journal

After every few experiments (or after significant findings), update the "What's Been Tried" section in autoresearch.md. Include:

• What worked and why
• What didn't work and why
• Dead ends to avoid
• Current best result and how it was achieved

7. Maintain Ideas Backlog

When you discover promising but deferred optimizations, append them as bullet points to autoresearch.ideas.md. Don't let good ideas get lost. Prune stale or tried entries.

8. Loop

Go back to step 1.

State Files Reference

JSONL Schema

Each line in autoresearch.jsonl is either a config header or an experiment result:

Config header (first line, or on re-init):

Experiment result:

ASI (Actionable Side Information)

Always record ASI with every experiment. At minimum: {"hypothesis": "what you tried"}. On discard/crash, also include rollback_reason and next_action_hint. Add any other key/value pairs that capture what you learned — dead ends, surprising findings, error details, bottlenecks.

ASI is the only structured memory that survives reverts. Without it, future iterations waste time re-discovering the same dead ends.

Confidence Scoring

After 3+ experiments, the helper script computes a confidence score using Median Absolute Deviation (MAD):

The score is advisory — it never auto-discards. If confidence is below 1.0x, consider re-running the same experiment to confirm before keeping.

Context Management

Droid sessions have finite context. To handle this gracefully:

1. Track experiment count in the current session. After ~15 experiments, context is getting heavy.
2. Save state proactively — all state lives in files (jsonl, md), so a new session can resume immediately.
3. When context is getting exhausted: update autoresearch.md with current findings, commit state files, and stop. The next session reads the files and continues.
4. On resume: read autoresearch.md, autoresearch.jsonl, and git log --oneline -20 to understand where things stand. Check current status:

Loop Rules Summary

• LOOP FOREVER. Never ask "should I continue?"
• Primary metric is king. Improved -> keep. Worse/equal -> discard.
• Annotate every run with ASI. Record what you learned, not just what you did.
• Watch the confidence score. < 1.0x means within noise — re-run to confirm.
• Simpler is better. Removing code for equal perf = keep.
• Don't thrash. Repeatedly reverting the same idea? Try something structurally different.
• Crashes: fix if trivial, otherwise log and move on.
• Think longer when stuck. Re-read source files, study the data, reason about what's actually happening. The best ideas come from deep understanding.
• Resuming: read autoresearch.md + git log, continue looping.

Finalization

When the experiment loop ends (termination condition met, user interrupts, or context exhausted), finalize the results into clean, reviewable branches. This is the last phase of an autoresearch session.

Step 1: Summarize Results

Review the git log for actual commits:

Step 2: Group Changes

Group kept experiments into logical changesets. Each group should:

• Represent a single coherent optimization or change
• Not share modified files with other groups (so branches can merge independently)
• Have a clear description of what it achieves and the metric improvement

Present the proposed grouping to the user for approval:

Wait for user confirmation before proceeding. In mission worker mode, proceed with the best grouping without waiting for confirmation.

Step 3: Resolve File Conflicts

If groups share files, resolve before creating branches:

• Merge the groups into one (if changes are related)
• Split the file changes more carefully (if they're truly independent modifications to different parts)
• Ask the user which group gets priority

Groups must not share files — each branch must be independently mergeable. If all changes touch the same file and can't be separated, create a single finalized branch with all improvements combined.

Step 4: Create Clean Branches

For each group:

Step 5: Verify and Report

For each finalized branch, run the benchmark to confirm the improvement holds, run any checks if applicable, and verify it merges cleanly with main.

Present a summary to the user:

The original experiment branch is always preserved — finalization creates new branches.

Mission Worker Mode

When running as a mission worker, the feature description specifies the optimization goal, termination condition, files in scope, and constraints. Read it carefully, follow the same loop procedure above, and respect the termination condition. When the condition is met, run finalization and report results in the handoff.