pi-autoresearch-loop

---
name: pi-autoresearch-loop
cli-support: [all]
description: Autonomous experiment loop for pi that continuously tries optimizations, measures results, and keeps what works
triggers:
  - autoresearch
  - autonomous experiment loop
  - optimize automatically
  - run experiment loop
  - continuous optimization
  - benchmark and improve
  - start autoresearch session
  - keep what works discard what doesnt
---

# pi-autoresearch — Autonomous Experiment Loop

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection

Autonomous experiment loop extension for [pi](https://github.com/antiwork/pi). Continuously proposes changes, benchmarks them, commits wins, reverts losses, and repeats — forever. Works for any measurable target: test speed, bundle size, build time, LLM training loss, Lighthouse scores.

---

## Installation

```bash
pi install https://github.com/davebcn87/pi-autoresearch

Manual install:

cp -r extensions/pi-autoresearch ~/.pi/agent/extensions/
cp -r skills/autoresearch-create ~/.pi/agent/skills/

Then /reload in pi.

---

Quick Start

/skill:autoresearch-create

The agent will:

1. Ask about your goal, command, metric, and files in scope (or infer from context)
2. Create a branch
3. Write autoresearch.md and autoresearch.sh
4. Run the baseline
5. Start looping immediately — no further input needed

---

Core Concepts

Two-file persistence model

Every session is fully recoverable from two files:

File	Purpose
autoresearch.jsonl	Append-only log — one JSON line per run (metric, status, commit, description)
autoresearch.md	Living document — objective, what's been tried, dead ends, key wins

A fresh agent with zero memory can read these two files and continue exactly where the previous session left off.

Session files written by the skill

File	Purpose
autoresearch.md	Session document — objective, metrics, files in scope, experiment history
autoresearch.sh	Benchmark script — pre-checks, runs the workload, outputs METRIC name=number lines
autoresearch.checks.sh	(optional) Backpressure checks — tests, types, lint. Failures block keep

---

Extension Tools

init_experiment

One-time session configuration. Call once at session start.

await init_experiment({
  name: "vitest-speed",
  metric: "seconds",
  unit: "s",
  direction: "lower", // "lower" | "higher"
});

run_experiment

Runs any shell command, times wall-clock duration, captures stdout/stderr.

const result = await run_experiment({
  command: "pnpm test --run",
  timeout_seconds: 120,           // optional, default 300
  checks_timeout_seconds: 300,    // optional, for checks script
});
// result: { exit_code, duration_seconds, stdout, stderr }

log_experiment

Records result, auto-commits on keep, updates the status widget and dashboard.

await log_experiment({
  metric_value: 42.3,
  status: "keep",          // "keep" | "discard" | "crash" | "checks_failed"
  description: "Enable parallel test workers in vitest config",
  commit_message: "perf: parallel vitest workers → 42.3s (-18%)",
});

---

The Autonomous Loop

Once started, the agent runs this cycle indefinitely:

propose change → edit files → run_experiment → measure metric
       ↓
  metric improved?
    YES → log_experiment(keep) → auto-commit → update autoresearch.md
    NO  → log_experiment(discard) → git revert → try next idea
       ↓
  repeat forever (until interrupted)

Interrupt anytime with Escape, then ask for a summary of what was tried.

---

Benchmark Script Format

autoresearch.sh must output at least one METRIC line:

#!/bin/bash
set -euo pipefail

# Pre-checks
[ -f package.json ] || { echo "No package.json"; exit 1; }