AgentEngineConfig.agent_framework Literal rejects "a2a" — out of sync with _SUPPORTED_AGENT_FRAMEWORKS and the GA A2A template

Start with vertexai/_genai/types/common.py and inspect the AgentEngineConfig, Dict twin, and update-config variants named in the report, comparing them with _SUPPORTED_AGENT_FRAMEWORKS in vertexai/_genai/_agent_engines_utils.py. Reproduce the shown agent_engines.create configuration, then run the relevant existing validation tests; done means agent_framework="a2a" is accepted consistently without weakening validation.

Environment

• google-cloud-aiplatform 1.159.0 → 1.165.1 (still present on main)
• GenAI client path: vertexai.Client().agent_engines

Bug

A2A on Agent Engine is a supported framework: _SUPPORTED_AGENT_FRAMEWORKS includes "a2a", the A2aAgent template sets agent_framework = "a2a" (since 1.147.0), the A2A framework template is GA (CHANGELOG 1.150.0), and the backend API accepts spec.agent_framework = "a2a".

But the config type was never updated to match:

# vertexai/_genai/types/common.py
agent_framework: Optional[
    Literal["google-adk", "langchain", "langgraph", "ag2", "llama-index", "custom"]
]

"a2a" is missing from the Literal (in AgentEngineConfig, its Dict twin, and the update-config variants), so passing it raises a Pydantic validation error, while the same package accepts it internally:

# vertexai/_genai/_agent_engines_utils.py (~line 167)
_SUPPORTED_AGENT_FRAMEWORKS = frozenset(
    ["google-adk", "langchain", "langgraph", "ag2", "llama-index", "custom", "a2a"]
)

Steps to reproduce

import vertexai
client = vertexai.Client(project="...", location="us-central1")
client.agent_engines.create(
    agent=my_agent,
    config={"staging_bucket": "gs://...", "agent_framework": "a2a"},
)
# pydantic ValidationError: Input should be 'google-adk', 'langchain',
# 'langgraph', 'ag2', 'llama-index' or 'custom'

Expected behavior

agent_framework="a2a" accepted in the config, consistent with _SUPPORTED_AGENT_FRAMEWORKS, the A2aAgent template, and the backend API.

Workaround

Set the fields on the agent object instead of the config: use the A2aAgent template, or set my_agent.agent_framework = "a2a" before create() — _get_agent_framework(agent) reads the attribute and bypasses the config Literal.

Related

#6267 (classic ReasoningEngine.create path missing agent_framework support) — this issue is about the GenAI client config type being out of sync.