jules-autoresearch

An iterative autoresearch loop skill for pi that uses Jules (Google's cloud coding agent) as the compute engine.

What it does

Instead of spending your own tokens on repetitive evaluation-and-refinement cycles, you delegate the heavy work to Jules. Pi orchestrates the loop; Jules runs the actual task, measures the metric, and suggests the next parameter set.

initial keywords / params
        │
        ▼
┌─────────────────┐       target met?
│  Jules runs     │──YES──────────────► done ✓
│  the task in    │
│  a cloud VM     │
└─────────────────┘
        │ NO
        ▼
  parse results + new params (Jules suggests them)
        │
        └──────────────► next iteration

Two modes

Research mode — Jules reasons about an open-ended task and suggests improved keywords each iteration.

Eval-harness mode — Jules runs a deterministic eval script (e.g. python eval_harness.py --params '...') against a labelled dataset, reports precision/recall/F1, and proposes better parameters. Zero qualitative guesswork.

Quick start

# 1. Set your Jules API key (from jules.google.com/settings)
export JULES_API_KEY="your-key-here"

# 2. Find your connected GitHub repo
python scripts/jules_client.py list-sources

# 3a. Research mode
python scripts/autoresearch.py \
  --source "sources/github-owner-repo" \
  --task "Find papers about: {keywords}" \
  --keywords "transformers, attention" \
  --target "5 papers with >50 citations from 2022+" \
  --metric-type qualitative

# 3b. Eval-harness mode (structured params + your eval script)
python scripts/autoresearch.py \
  --source "sources/github-owner-repo" \
  --eval-script "scripts/eval_harness.py" \
  --params '{"queries": ["conference"], "exclusions": ["conference call"], "patterns": ["will present at"]}' \
  --target "precision >= 0.90 with recall >= 0.70" \
  --metric-type numeric --target-value 0.90 \
  --parallel 3

Files

├── SKILL.md                      # Full skill instructions for pi
├── scripts/
│   ├── jules_client.py           # Jules REST API wrapper (pure stdlib)
│   └── autoresearch.py           # Autoresearch loop orchestrator
└── references/
    ├── api_quickref.md           # Jules API quick reference
    └── eval_harness_template.py  # Copy-paste starting point for your eval script

Requirements

• Python 3.8+, no external dependencies (pure stdlib)
• A Jules account with API key (jules.google.com)
• A GitHub repo connected to Jules

Notes

• Jules sessions take 15–20 min to complete (most of that is Jules' internal state transitions, not actual work time)
• sendMessage requires OAuth — if Jules asks a question mid-session, the polling script surfaces the question and session URL so you can respond via the Jules web UI
• --parallel N in eval-harness mode means Jules evaluates N parameter variants in a single session (quota-efficient); in research mode it dispatches N separate sessions

As a pi skill

Drop the jules-autoresearch/ directory into ~/.agents/skills/ and pi will load it automatically.