Multi-Agent Autoresearch

This repo is a self-contained Open Source AI Lab that researches papers, manages experiments, runs GPUs, and repeats. It is based on autoresearch by Andrej Karpathy.

OpenCode is the primary way to use this repo. The checked-in agents and AGENTS.md tell the agent which scripts to run and how to use them safely. There are also Hermes and Pi Agent adapters, plus older Codex and Claude Code material.

What This Repo Contains

• train.py The working experiment surface.
• prepare.py Read-only benchmark setup and evaluation logic.
• research/results.tsv Append-only local run ledger.
• research/live/ The current promoted local master snapshot and DAG.
• research/ Notes, campaign state, experiment records, and templates.
• .opencode/agent/ The checked-in OpenCode agents: autolab, planner, experiment-worker, reviewer, memory-keeper, researcher, and reporter.
• .pi/ Optional Pi Agent project settings, role agents, and the /autolab prompt.
• .agents/skills/ Shared repo-local skills.

Local Setup

Install the repo and create your local operator env:

uv sync
hf auth login
hf auth whoami
opencode auth login
# optional for Hermes:
hermes setup
# optional for Pi Agent:
npm install -g @mariozechner/pi-coding-agent

If you have not warmed the shared Hugging Face cache yet, you can ask OpenCode to do that as part of the first session. The exact script path is already in AGENTS.md.

Start Hermes

From the pre-training project root:

uv run scripts/setup_hermes_profile.py --profile autolab
uv run scripts/print_hermes_kickoff.py --gpu-slots 1
autolab chat --toolsets "terminal,file,web,skills,delegation,clarify"

Hermes loads AGENTS.md automatically, so the repo intentionally does not ship .hermes.md.

Use the parent Hermes session to delegate planner, reviewer, researcher, reporter, experiment-worker, and memory-keeper roles. Keep Hermes child concurrency at 3 or fewer per parent session. Use uv run scripts/hermes_worker.py create ... plus