Jack Wotherspoon, a developer experience engineer at Google DeepMind, demonstrates Gemini Managed Agents in this twenty-minute AI Engineer presentation.
Move the runtime off the laptop
- Managed agents package a cloud sandbox and a harness, the software that runs the model’s tools, files and commands.
- Wotherspoon describes a Gemini API call that starts a remote environment with Google’s Antigravity harness.
- Load a repository, skills and
AGENTS.mdinstructions into that environment to reproduce a team’s workflow. - Built-in execution and file tools can be extended through custom functions and MCP, a standard for connecting agents to tools.
Separate conversation state from files
- An interaction holds conversation history; an environment holds files and other workspace artifacts.
- Give a new agent the environment ID to reuse files while starting with fresh conversational context.
- In the demo, an SEC filing researcher saves company metrics, a quant agent analyzes them and a third agent prepares a report.
- Save an agent configuration for repeat use, and use scheduled triggers for recurring tasks.
Keep credentials outside the workspace
- Wotherspoon describes a proxy that replaces dummy credentials with real secrets for matching outgoing requests.
- This reduces direct secret exposure inside the sandbox; the permitted API calls still need careful scoping.
- Step traces expose tool activity so developers can inspect and revise the workflow.
Read the demos as architecture examples
- Additional examples put managed agents behind a receipt-reporting app and a stock-research interface in AI Studio.
- Narrowly defined agents and file handoffs make each stage easier to inspect than a single broad assignment.
- Caveat: this is a Google product demonstration. A generated stock forecast or sourced PDF does not validate financial accuracy or prove the credential design prevents every misuse.
“It has a very fresh kind of context window and conversation, but it’s reusing the environment.” — Jack Wotherspoon, 10:16–10:20