From one agent to a governed AI org

Molecule AI uses an agent-to-agent (A2A) protocol layered on JSON-RPC 2.0. Workspaces discover each other through the platform registry and exchange messages peer-to-peer over A2A, so the control plane is never in the message path itself. The Model Context Protocol (MCP) sits next to A2A and exposes the platform's tools (delegation, memory, approval, sandbox) to every agent — agents call platform capabilities the same way they call peer agents, so teams don't have to maintain two integration paths (one for tool calls, another for peer delegation). Both synchronous (delegate-and-wait) and asynchronous (fire-and-poll) delegation are supported, so an agent can either wait inline on a small follow-up or kick off a long-running task and come back later. Every message, every tool call, and every approval is logged in the audit trail. The benefit is a single composable surface: one consistent way to send work, one consistent way to see what happened.

Each workspace in Molecule AI picks the runtime that drives its single agent — model, system prompt, and tools are supplied at provision time, so every runtime works out of the box with no per-agent code. Today's adapter set includes Claude Code, Codex, Hermes (sidecar), OpenClaw, and Google ADK as a first-class runtime in development. The hosted SaaS ships nine adapters across the open-core monorepo. New adapters can be added by following the adapter authoring guide in the documentation; the platform-side surface an adapter must implement is bounded — start, message, stop, status, capabilities — so the integration cost is small and predictable. The control plane is deliberately runtime-neutral: it does not favor any one SDK, and the canvas treats a Claude Code agent and a Gemini-on-ADK agent as the same kind of object.

Governance and observability are first-class layers of the Molecule control plane, not bolted on after the fact. The platform's audit trail is the integration: every A2A message, every task update, every skill promotion is appended to a JSON Lines log by default, so teams don't have to wire up a separate logging stack to see what their agents did. Tracing covers every LLM call end-to-end with Langfuse + OpenTelemetry, including tokens and cost reported per call. If a runtime adapter misbehaves, the audit trail still records the boundary call, so failures stay diagnosable from outside the agent. The Governance & Topology layer at the top of this page summarizes the public surface; deeper architecture details live in the developer-facing documentation.

Most multi-agent platforms ask teams to maintain two pictures of their work in parallel: a visual diagram for planning and stakeholder review, and a separate configuration model for what's actually running. The two drift over time, and the drift gets expensive — engineers update one and forget the other, product managers look at a stale picture, new hires can't tell which view is true. Molecule AI is built around the idea that the org chart on the canvas is the working representation of the team. Teams sketch how their agents should collaborate the same way they'd sketch a human org chart: lead at the top, specialists reporting in, peers grouped where they share work. The benefit is design-time speed and a shared mental model — engineers, product managers, and team leads all work from the same canvas as the team grows and reshapes, instead of translating between a whiteboard, a config file, and a runtime view.