autoresearch.md

name	autoresearch
description	Autonomous optimization loop. The agent modifies a file, a benchmark measures the result, an evaluation gate keeps or reverts. Repeats up to 50 times. Use when you want iterative, overnight optimization with deterministic evaluation — the agent proposes changes, scripts decide outcomes.
parameters	name description target File the agent modifies, e.g. 'optimize.js' or 'src/model.py'. name description goal What to optimize: 'execution speed', 'memory usage', 'accuracy', etc. name description benchmark Shell command that benchmarks the target. Must exit 0 on success, 1 on crash. Should write metrics to a file the evaluate command can read. name description evaluate Shell command that compares the benchmark result to the previous best. Exit 0 if improved (and checkpoint the change), exit 1 if not (and revert the target file). name description recover Shell command to restore the target file after a crash, e.g. 'cp optimize.best.js optimize.js'. name description max_experiments Maximum number of experiments to run, e.g. '50' for overnight or '10' for a quick test.

task: "Optimize {{target}} for {{goal}}" success_criteria: "The optimization loop runs until max_runs is reached. Improvements are kept, non-improvements are reverted by the evaluate script." entry_step: experiment artifacts:

• name: experiment_log description: "Accumulated log of what each experiment tried and why. Each entry is appended automatically — the actor writes one entry per iteration." steps:

• type: action name: experiment actor: worker instruction: | You are an autonomous researcher optimizing {{target}} for {{goal}}.

  Read {{target}} to understand the current implementation.

  Read the experiment_log artifact to see what has been tried before. Do NOT repeat approaches that have already been tried. If an approach was tried and discarded, try something structurally different.

  Propose and apply ONE optimization to {{target}}. Be creative but disciplined — try one idea at a time. Consider:

  • Algorithmic improvements
  • Data structure changes
  • Memory access patterns
  • Eliminating redundant work
  • Mathematical shortcuts

  Do NOT modify any other files. The benchmark and evaluation scripts are fixed.

  Write an entry to experiment_log describing what you tried and why: { "approach": "what you changed", "rationale": "why this should help" } reads: [experiment_log] writes: [experiment_log] routes: { done: benchmark } max_runs: "{{max_experiments}}"

• type: command name: benchmark command: "{{benchmark}}" on_success: evaluate on_failure: recover

• type: command name: evaluate command: "{{evaluate}}" on_success: experiment on_failure: experiment

• type: command name: recover command: "{{recover}}" on_success: experiment on_failure: failed

• type: terminal name: failed outcome: failure summary: "Recovery failed after a crash — manual intervention needed."