Agent-to-Agent (A2A) protocol

Dwara can act as a gateway for the Agent-to-Agent (A2A) protocol, an emerging standard for machine-to-machine communication between autonomous AI agents. In the A2A model, an agent publishes an Agent Card describing its identity and capabilities, and other agents submit tasks to it as JSON-RPC messages over HTTP. In Dwara, each remote A2A agent is registered as an AI provider: a chat request routed to it is translated into an A2A task-submit call, so clients keep speaking the OpenAI-shaped chat API while the peer speaks A2A.

Status: partial. The provider adapter (routing, task-submit translation, Agent Card parsing) is wired; the A2A task lifecycle -- long-running task negotiation, cancellation, and status endpoints -- is stubbed pending spec freeze. The block is feature-gated behind the a2a cargo feature: without it, config is accepted but inert (validation warns). There is no /.well-known/agent-card.json discovery endpoint in this build.

When to use this

Register A2A agents when a fleet of agents must call each other through a controlled boundary -- one ingress with the gateway's authn/authz, rate limits, token budgets, and analytics applied to agent-to-agent calls, instead of agents dialing each other directly.

If your agents communicate over a single fixed channel with no policy boundary, an A2A hop adds latency for no benefit.

Configuration

A2A agents live under ai.a2a. Each agent names a base URL (the A2A endpoint) and an Agent Card (a file path or inline JSON, parsed at compile time). The agent then appears as a provider of kind a2a, and a model alias routes to it by name:

ai:
  a2a:
    enabled: true
    agents:
      - name: mock-agent
        url: http://agent-upstream:8080
        card:
          name: mock-agent
          description: demo A2A agent
          version: "1.0"
  providers:
    - name: agent-pool
      kind: a2a
      agent: mock-agent
  models:
    - alias: a2a-echo
      provider: agent-pool

A chat request with "model": "a2a-echo" is translated into an A2A task-submit body and placed through the agent's URL; the response is mapped back to the chat-completion shape (content, finish_reason, usage). Authn, authz, rate limits, and the rest of the policy chain apply to the request exactly as they would to any other route -- agent-to-agent calls present the same credentials any client would.

Field	Default	Description
enabled	false	Master switch: the block is inert until enabled, so config can be staged ahead of activation.
agents[].name	--	Unique agent name; referenced by the provider's agent field.
agents[].url	--	The agent's base URL (http/https); the adapter builds the task-submit path under it.
agents[].card	--	Agent Card: a filesystem path to JSON or an inline JSON object; parsed at compile time.
sessions	built-in defaults	Session policy (TTL 3600 s, max 1000 concurrent), mirroring the MCP gateway.

Notes

• A2A is an emerging standard; the card format and JSON-RPC method names track the current draft and may change.
• The Agent Card is a static document (file or inline) parsed at compile time. The gateway does not serve cards over HTTP in this build.
• The gateway does not itself execute agent tasks -- it translates and routes the messages. Long-running task state and streaming are handled by the agents at either end; the lifecycle surface is stubbed pending spec freeze.

Runnable demo

demos/07-ai-gateway/ (test script: test-15-a2a.sh) in the repository routes a chat request through an ai.a2a agent to a mock A2A upstream and asserts the full chat-to-task-submit translation, including the task-completion state and usage accounting. The category README covers prerequisites and teardown.