Anthropic Claude model on Vertex fails when served as an A2aAgent on Agent Engine (serving instance killed mid-request → 400 on message:send); Gemini works

[BrianAlderman]

Existing issue check

Closest are #3861 (AdkApp↔A2aAgent comms, Gemini + Sequential agent — closed stale, not resolved) and #3020 (Vertex "model not found" — closed). Neither covers this Claude-specific A2A serving failure; searches for Anthropic/Claude + A2A return nothing.

Describe the bug

An ADK LlmAgent whose model is an Anthropic Claude model on Vertex (via google.adk.models.anthropic_llm.Claude / AnthropicLlm), deployed to Vertex AI Agent Engine as a managed A2aAgent (vertexai.preview.reasoning_engines.A2aAgent + agent_engines.create), fails to serve over A2A. On an a2a/v1/message:send call the serving instance is killed mid-request (clean SIGTERM + instance replacement) and the caller receives HTTP 400.

The same agent/topology with a Gemini model serves cleanly over the identical A2A path. The failure tracks the model, not tools/MCP.

Environment

• google-adk: 2.3.0 (2.4.0 changelog shows no related fix)
• a2a-sdk: 0.3.26 ([http-server] extra installed)
• google-cloud-aiplatform: 1.159.x
• Runtime: Vertex AI Agent Engine (managed A2aAgent)
• Model: Anthropic Claude on Vertex (publishers/anthropic/models/…), region global
• Client: ADK RemoteA2aAgent resolving the agent card (…/a2a/v1/card)

To reproduce

1. Build a minimal LlmAgent whose model is an Anthropic Claude-on-Vertex model (ADK Claude(...)), with one trivial tool or none.
2. Wrap it as A2aAgent(agent_card=…, agent_executor_builder=…) and deploy via agent_engines.create(agent_engine=A2aAgent(...), requirements=[…, "a2a-sdk[http-server]"], …).
3. Wait for the engine to stabilize; confirm the card at …/a2a/v1/card (HTTP 200).
4. Send one message via a2a/v1/message:send (e.g. through ADK RemoteA2aAgent).

Observed behavior

• The task is created, then the serving instance logs a clean SIGTERM and a replacement container starts mid-request (~10–45s in), around/before the model turn.
• The client's message:send returns HTTP 400; upstream this surfaces as an empty result.
• It fails even on a no-tool ping — before any tool call, no MCP involved.
• The Claude call itself works: …:rawPredict returns 200 in the logs. So inference is fine — the break is in the native A2A serving / response-conversion path, not the model call.

Isolation

Agent	Model	Tools/MCP	A2A serve
control	Gemini	none	PASS (clean reply, no SIGTERM)
minimal	Claude (Sonnet)	one trivial tool, no MCP	FAIL (SIGTERM → 400)

The only variable that tracks the failure is the model (Claude vs Gemini). A heavier Claude agent (many tools + MCP) fails the same way, and hours after deploy (ruling out a fresh-deploy confound).

Expected behavior

A Claude-on-Vertex LlmAgent served as an A2aAgent should handle message:send and return a response, as the Gemini equivalent does.

Hypothesis / notes

• Likely the Claude event/part structure (e.g. thinking blocks or streaming aggregation) differs from Gemini in a way the A2A message/event converter (google/adk/a2a/converters/…) or the A2aAgentExecutor response path doesn't handle, taking down the serving worker.
• Untested cheap idea: disable Claude "thinking" (thinking_budget=0) so it emits plain text/tool_use like Gemini.
• Happy to provide a fully self-contained minimal repro + engine logs on request.