autoresearch-coordinator

Autoresearch Coordinator

Coordinate distributed autoresearch experiments across GPU workers discovered on-chain via ERC-8004 and paid per-experiment via x402 micropayments. This replaces the Ensue-based shared-memory coordinator from autoresearch-at-home with a fully decentralised discovery and payment loop built on obol-stack primitives.

When to Use

• Discovering GPU workers advertising devops_mlops/model_versioning capabilities via the 8004scan public index API
• Probing worker endpoints for x402 pricing before submitting experiments
• Submitting train.py experiments to remote GPU workers through x402 payment gates
• Running the continuous THINK/CLAIM/RUN/PUBLISH experiment loop
• Viewing the global leaderboard of autoresearch results from worker metadata
• Coordinating multi-worker experiment campaigns

When NOT to Use

• Selling your own GPU as a worker -- use autoresearch-worker (then monetize it with obol sell http)
• Buying generic inference (chat completions) -- use buy-x402
• Discovering agents without running experiments -- use discovery
• Signing transactions directly -- use ethereum-local-wallet
• Cluster diagnostics -- use obol-stack

Quick Start

# Discover available GPU workers from the preferred public index API
python3 scripts/coordinate.py discover

# Discover with custom limit
python3 scripts/coordinate.py discover --limit 5

# Probe a specific worker for pricing
python3 scripts/coordinate.py probe https://worker.example.com/services/autoresearch-worker

# Submit a single experiment to a worker
python3 scripts/coordinate.py submit https://worker.example.com/services/autoresearch-worker train.py

# Submit with custom config overrides
python3 scripts/coordinate.py submit https://worker.example.com/services/autoresearch-worker train.py \
  --config '{"batch_size": 64, "learning_rate": 0.001}'

# View global leaderboard (best val_bpb across all workers)
python3 scripts/coordinate.py leaderboard

# Run continuous experiment loop (discover -> pick -> submit -> publish)
python3 scripts/coordinate.py loop train.py

# Loop with worker preference and max rounds
python3 scripts/coordinate.py loop train.py --prefer https://worker.example.com/services/autoresearch-worker --rounds 10