AgentReviewers

AI judge swarm that reviews your code and stores every verdict on 0G Storage. (

AgentReviewers

Created At

Open Agents

Project Description

Autonomous Agent Reviewers (AAR) is a swarm of AI judge agents that peer-review hackathon submissions, coordinating entirely through 0G Storage on the 0G Galileo testnet. When a project is submitted, an intake agent fetches the repository metadata, uploads the submission record to 0G, and fans out to a three-judge panel — one tuned for technical merit, one for originality, and one trained as a deliberate skeptic. Each judge reads the submission from storage, casts an independent verdict via Anthropic's Claude using structured tool-use, and writes its scored reasoning back to 0G. Every payload that crosses an agent boundary is content-addressed by its 0G root hash and validated against a shared zod schema on both ends, so the on-chain record is the source of truth — no agent ever trusts another agent's HTTP body for anything beyond a hash.

After the first round, an aggregator agent runs a single round of deliberation: each judge sees its peers' verdicts and either revises its score or holds firm, with both outcomes recorded as distinct on-chain artifacts. The aggregator then computes a weighted final score, flags meaningful disagreement, and produces a panel verdict that captures the full deliberation trail — round-1 verdicts, revisions, abstentions, and a dissent summary when the panel didn't converge. Each agent runs as an isolated Express service with its own funded wallet, so concurrent writes to 0G don't collide on nonces, and a live dashboard streams the agents' JSONL logs over Server-Sent Events so you can watch the panel reason through a submission in real time. The result is a fully auditable, reproducible review pipeline where every score, every change of mind, and every disagreement is provably stored on decentralized infrastructure.

How it's Made

AAR is built as a small fleet of Node/Express services that talk to each other only in 0G Storage root hashes — never raw payloads. The five agents (intake, three judges, aggregator) each live in their own folder with their own package.json and node_modules, sharing code through an aar-shared module linked via pnpm's link: protocol so edits to common modules propagate without reinstalls. A separate log-streamer service tails every agent's pino JSONL log with chokidar (in polling mode — see hacks below) and broadcasts entries over Server-Sent Events to a Vite + React 19 + TypeScript dashboard. The dashboard is a passive consumer: it derives every visualization from the canonical event vocabulary the agents already emit, and adds zero new endpoints or log events of its own.

background image mobile

Join the mailing list

Get the latest news and updates

AgentReviewers | ETHGlobal