Molecules AI
The operating system for AI agent organizations.
Overview
Molecules AI is the platform layer for the multi-agent era. Instead of wiring agents together with brittle workflow logic, you model them as an organization — roles, reporting lines, memory scopes, and approval ladders — and the control plane handles coordination.
I'm the founder, and the bet is simple: the unit of intelligence isn't the model, it's the org. As teams go from one agent to dozens, the bottleneck stops being the model and becomes how agents coordinate, who can approve what, and how knowledge accumulates across runs.
The product is open-core. The Canvas runs in your browser as a visual topology editor with a live A2A trace overlay; the Control Plane runs on your own tenant with a JSON-RPC 2.0 A2A router, a three-tier memory store, and enterprise governance. Bring whatever runtime you already trust — Claude Code, LangGraph, CrewAI, AutoGen — the topology never changes.
It won 1st place at the NVIDIA DGX Spark Hackathon (China, 2026).
Key features
Role > Task
Model the org chart, not the workflow. Roles inherit tools, memory scope, and an HITL ladder.
Distributed A2A protocol
Agents speak JSON-RPC 2.0 directly. The platform routes; it's never a single bottleneck.
Three-tier memory
LOCAL (what an agent saw), TEAM (what the squad agreed), GLOBAL (what the org learned) — one continuous, versioned memory.
Runtime-portable
Swap runtimes — Claude Code, LangGraph, CrewAI, AutoGen — without touching the topology.
Governance built in
Tiered HITL approvals, audit trails, RBAC, AES-256-GCM encryption, and EU AI Act readiness.
Three-layer marketplace
Buy or sell a plugin, hire a pre-trained agent, or import an entire department as a signed bundle.