A2A in Cognition

Cognition implements A2A 1.0 as a protocol surface over the same durable agent runtime used by its native APIs. A2A does not create a second execution engine: Agent Cards, JSON-RPC requests, tasks, streaming events, and artifacts project onto Cognition's existing agent, session, run, scope, and storage primitives.

Implementation model

A2A caller
  -> builder-owned gateway authentication and authorization
    -> trusted effective-scope headers
      -> Cognition A2A JSON-RPC adapter
        -> durable task and session runtime
          -> model, tools, sandbox, artifacts, and observability

Cognition owns protocol validation, Agent Card generation, durable execution, streaming, persistence, and exact-scope isolation. Builders own identity, authorization, route selection, and the mapping from authenticated claims to trusted scope headers.

Concepts

Concept	What it explains
Agent Cards and public skills	Public identity, endpoints, MIME modes, skills, and the boundary between discovery and runtime configuration.
Message Parts and artifacts	Text, structured data, inline bytes, URL references, persistence, and execution boundaries.
Tasks and streaming	Durable task identity, runs, contexts, continuation, cancellation, and event delivery.
Security and scoping	Trusted ingress, exact-scope isolation, authentication discovery, and builder responsibilities.

For setup and invocation steps, use the A2A Builder Guide. For exact endpoints and request bodies, use the API Reference.

Supported surface

Cognition advertises the JSONRPC binding and A2A protocol version 1.0. It supports SendMessage, SendStreamingMessage, GetTask, ListTasks, CancelTask, and SubscribeToTask. Streaming is enabled. Push notifications, gRPC, HTTP+JSON, and authenticated extended Agent Cards are not currently exposed.