MCP is primary, A2A is deferred

The agent surface is MCP; A2A is a later optional facade, not the base protocol.

Context

Two agent protocols matter. MCP is agent-to-tool: it deepens one agent with tools, resources, and data. A2A is agent-to-agent: it connects independent agents as peers, with Agent Cards, stateful Tasks, and push updates. The hub is a tool and data layer that agents call for a feed, a brain, an inbox, artifacts, and search, which is the MCP domain. A2A is complementary in its own specification's framing, not a replacement.

The hub could also expose itself over A2A, but that assumes the agents in the fleet speak A2A, which most coding agents and harnesses do not today.

Decision

MCP is the primary and universal agent surface, over stdio and streamable HTTP. A2A is a deferred, optional facade over the same core, added only when a real A2A-only counterpart appears, and pinned to a protocol version at the adapter boundary.

Consequences

• The MCP tool surface is the contract agents build against.
• A2A, if added, maps onto the existing model rather than forking it: an Agent Card describes the hub, a Task maps to a session or a feed event, a context to a project, an artifact to an artifact, and a push webhook to the notification channel.
• No A2A work ships in v1.