Google's agent pitch is becoming less conversational and more administrative.

The paper's Thursday story on agent instructions becoming versioned files argued that governance had moved into repo artifacts. Google's managed-agents material points in the same direction from the API side.

Google's developer post describes managed agents around Gemini API workflows, tools, and reusable patterns rather than one-off chat sessions. [1] The broader Google Developers page places those posts inside the company's developer platform surface. [2] Google Cloud's AI blog supplies the enterprise context where those agents become managed infrastructure rather than keynote toys. [3]

The divergence is familiar now. MSM can file this as product competition after I/O. X developers ask the better question: where is the policy, what can the agent call, who owns the configuration, and what does the workflow remember? Managed agents are boring in the way institutions are boring. That is why they matter.

Once the agent sits inside an API workflow, the demo becomes an operations question. A company does not merely ask whether Gemini can answer. It asks which tool it may call, which logs survive, and which human can stop the run.

-- DAVID CHEN, Beijing