Skip to content

RESULTS.md

Latest commit

 

History

History
57 lines (42 loc) · 2.42 KB

RESULTS.md

File metadata and controls

57 lines (42 loc) · 2.42 KB

Experiment: subagents == agent-as-tool composition (confirmed)

The capability map (src/agentlift/capabilities.py) classifies "subagents -> coordinator" as native (Anthropic) or emulated (Google, OpenAI) -- never unsupported. This experiment is the receipt: the coordinator-delegates-to-a-sub-agent pattern runs on both Google and OpenAI today.

Both scripts read keys from the environment and run the orchestration locally -- which is the point: model calls go to the provider, but the routing between agents runs in your process. No deploy required.

OpenAI (Agents SDK, researcher.as_tool())

$ python openai_agent_as_tool.py
QUESTION: How tall is the Eiffel Tower in meters, and what year was it completed?

FINAL ANSWER:
  The Eiffel Tower is 324 meters tall (including antennas) and was completed in 1889.

--- delegation trace (proof the coordinator called the sub-agent as a tool) ---
  function_call  ask_researcher
  ToolCallOutputItem

The coordinator called the researcher sub-agent as a tool (ask_researcher), got the result, and synthesized the answer. Agent-as-tool == subagent. The loop ran in-process.

Google (ADK, sub_agents=[...])

$ GOOGLE_GENAI_USE_VERTEXAI=FALSE python google_adk_subagents.py
QUESTION: How tall is the Eiffel Tower in meters, and what year was it completed?
--- event trace ---
  [delegation] coordinator -> transfer_to_agent({'agent_name': 'researcher'})
  [researcher] The Eiffel Tower is 330 meters tall.

The coordinator delegated to the researcher sub-agent via ADK's native transfer_to_agent. Deployed to Vertex Agent Engine, this delegation runs server-side as one reasoningEngine.

Conclusion

A subagent roster is a universal capability, not a per-provider lottery:

Provider Mechanism Where the delegation loop runs Tier
Anthropic native multiagent coordinator provider runtime (hosted) native
Google ADK sub_agents -> Agent Engine provider runtime (hosted, one resource) emulated
OpenAI agent.as_tool() / Agent Builder nodes your orchestrator (self-hosted) emulated

The difference is only where the delegation loop runs -- the provider's runtime, or yours. That is the precise meaning of "own the definition, rent the runtime": the roster is portable; the orchestration is either rented (Anthropic, Google) or self-hosted (OpenAI).

Confirmed 2026-06-03 with openai-agents + google-adk 2.1.0.