Research
SHIPPED
ĀTAVI
Multi-Agent Research Protocol
What It Is
A host-agnostic refinement protocol for multi-agent AI research — five specialist agent roles, strict file-first workflow, anti-coherence-trap mechanisms, and fully auditable experiment records. Published on npm and actively used in production.
The Problem
Research agents hallucinate citations, converge on wrong answers, and can't be audited. ĀTAVI is the protocol layer that makes multi-agent research reliable — enforcing structure, novelty checks, and institutional memory across sessions.
Approach
- Five specialist agents (Theorist, Experimentalist, Scout, Critic, Synthesist) with formal hand-off contracts
- File-first .atavi/ workspace — all state is on disk, enabling resume and full audit trail
- Registries for claims, experiments, and prior art with cross-pollination mechanisms
- Institutional memory across sessions; convergence history tracked explicitly
- Host-agnostic design — starter prompts for Claude, Gemini, Codex, and local models
Results
- Shipped v0.1.0 — published to npm (npm install -g atavi)
- In production use for structuring AI-assisted research sessions
- GitHub: github.com/Murai-Labs/Atavi — MIT licensed, open protocol
Stack
Node.jsJavaScriptCLIMulti-agent orchestrationnpm