autoloop

name	autoloop
description	Universal autonomous optimization loop. Use this skill when the user wants to: continuously optimize any metric (latency, bundle size, test coverage, benchmark scores, build time, prompt quality, algorithm speed), run overnight experiments, set up an autonomous improvement loop on any codebase, or apply the autoresearch pattern to a new domain. Also trigger when the user mentions "autoloop", "optimize in a loop", "run experiments overnight", "fitness function", "keep or discard", "autonomous optimization", or references Karpathy's autoresearch pattern for non-ML workloads.
argument-hint	[recipe name or custom description]

Autoloop — Universal Autonomous Optimization

You are running an autonomous optimization loop — the generalized form of Karpathy's autoresearch. The pattern is simple and universal:

1. Define genome files — the files you are allowed to modify
2. Define a fitness command — a command that produces a measurable score
3. Loop forever: mutate the genome, measure fitness, keep improvements, discard regressions

This works for ANY optimization problem that has a numeric score.

Quick Start

When the user invokes /autoloop, follow this protocol:

1. Detect or Ask

Determine what to optimize. Check for signals:

• Is there a autoloop.config.json in the project root? If so, load it.
• Does the project have benchmarks, tests, or build scripts? Suggest a recipe.
• Did the user specify a recipe name? Load it from references/recipes/.
• Otherwise, ask the user three questions:
  1. What files can I modify? (genome)
  2. What command measures success? (fitness)
  3. Should I minimize or maximize the score? (direction)

2. Configure

Create or update autoloop.config.json in the project root. See the schema below.

3. Branch

git checkout -b autoloop/<tag>

Use a tag based on today's date and the optimization target (e.g., autoloop/mar21-latency).

4. Baseline

Run the fitness command once without modifications. Record the result as the baseline in the results log.

5. Loop

Hand off to the loop-runner agent. It runs autonomously and never stops.

Config Schema

{
  "name": "string — human-readable name for this optimization run",
  "genome": {
    "files": ["glob patterns for files the agent may modify"],
    "constraints": ["natural language rules the agent must follow"]
  },
  "fitness": {
    "command": "shell command that produces output containing the metric",
    "metric_pattern": "regex with one capture group extracting the numeric score",
    "direction": "minimize | maximize",
    "timeout_seconds": 300
  },
  "budget": {
    "time_per_experiment": 300,
    "max_experiments": 0
  },
  "logging": {
    "file": "autoloop-results.tsv",
    "columns": ["commit", "score", "status", "description"]
  },
  "guardrails": {
    "max_score_regression": 0.0,
    "required_tests": "shell command that must pass before fitness is measured",
    "max_files_changed": 10
  }
}

Field Reference

Field	Required	Description
name	Yes	Descriptive name for the run
genome.files	Yes	Glob patterns for modifiable files. Everything else is off-limits.
genome.constraints	No	Natural-language rules (e.g., "Do not change the public API")
fitness.command	Yes	Command whose stdout/stderr contains the metric
fitness.metric_pattern	Yes	Regex with one capture group () that extracts the numeric value
fitness.direction	Yes	"minimize" (lower is better) or "maximize" (higher is better)
fitness.timeout_seconds	No	Kill the fitness command after this many seconds (default: 300)
budget.time_per_experiment	No	Expected wall-clock seconds per experiment (for estimation)
budget.max_experiments	No	Stop after N experiments. 0 = infinite (default)
logging.file	No	Path to TSV results log (default: autoloop-results.tsv)
logging.columns	No	Column names for the TSV (default shown above)
guardrails.required_tests	No	Command that must exit 0 before fitness is measured
guardrails.max_score_regression	No	Maximum allowed regression from best score (for multi-metric)
guardrails.max_files_changed	No	Refuse to run if more than N files are modified in one experiment

Recipes

Pre-built configurations for common domains. Load with /autoloop <recipe>.

Recipe	Optimizes	Typical Fitness
web-perf	Lighthouse / Core Web Vitals scores	npx lighthouse --output json
api-latency	HTTP endpoint response time	wrk or autocannon benchmarks
test-coverage	Code coverage percentage	jest --coverage, pytest --cov
prompt-engineering	LLM output quality via eval harness	Custom eval script
build-optimization	Build time or artifact size	time npm run build
algorithm-speed	Benchmark execution time	hyperfine, pytest-benchmark
bundle-size	JavaScript bundle size in bytes	size-limit or webpack stats
security-scan	Security finding count	npm audit, bandit, semgrep
custom	Anything — user defines everything	User-provided

Multi-Metric Strategy

Sometimes you want to optimize one metric while keeping others from regressing. Use guardrails.required_tests for hard constraints and track secondary metrics in the log.

See references/strategies.md for advanced patterns:

• Pareto optimization (multiple fitness functions)
• Staged optimization (optimize A first, then B)
• Constrained optimization (optimize A subject to B > threshold)

The Autoresearch Lineage

This plugin generalizes the pattern from Karpathy's autoresearch:

Autoresearch	Autoloop
train.py	Any genome files
val_bpb	Any fitness metric
uv run train.py	Any fitness command
ML hyperparameters	Any code modifications
GPU training	Any compute workload

The core loop is identical. The abstraction makes it work everywhere.