Google DeepMind’s Philipp Schmid says its managed agent environments let developers give an AI agent a persistent place to run code, install tools and continue work across requests. The practical pitch is fewer bits of infrastructure for developers to assemble themselves, and a shift from chatting with a model towards setting it a goal and checking what it does.
Google DeepMind Watch analysis
What happened
In an account of an AI Engineer podcast conversation published by BigGo Finance on 7 October, Schmid described Google’s Interactions API and agent stack. He said developers can call a remote environment where an agent can execute code, create files, install dependencies and connect to developer APIs. The environment can persist across interactions, and agents can share one, allowing one agent or sub-agent to pick up work another has done.
Schmid described a synchronous timeline of typed steps in the Interactions API, so a function result is represented as a function result rather than being folded into a user message. He characterised the API as Google’s answer to OpenAI’s Responses API, making a claim about developer experience, not model quality. BigGo also reports that he described a Gemini API CLI for agents to call the API and build or improve agents, distinct from Gemini CLI and from a programming agent.
Why it matters
The shift is from developers managing each function call and its execution environment to asking an agent to work inside a managed, reusable space. That could reduce setup for teams building agents, while making persistent environments and shared context central parts of how those systems are distributed.
The same features raise practical questions about what persists, who can inspect a shared environment and how its contents are controlled. BigGo’s account says the environments are both managed and persistent, but does not spell out the details of data retention or cleanup. Those are not footnotes for anyone putting real work in the sandbox.
Our read
This is a more consequential pitch than another model benchmark: Google is trying to make the agent’s workspace part of the API. Reusing an environment could save developers from infrastructure chores, but the useful test is whether teams can understand and control what carries over. Before building on it, get clear answers on persistence, access and cleanup. A tidy API is lovely; a tidy data boundary is lovelier.
What to watch
- How Google documents environment persistence, data retention and cleanup.
- Whether developers can inspect and control what agents share across environments.
- How the Interactions API and managed environments compare in practice with competing agent APIs.
Discussion spark: Would you trust a managed, persistent sandbox to make agent development simpler, or does sharing state across agents create too much uncertainty about control and data?
Sources and evidence
- Google DeepMind's Philipp Schmid: AI Agents Should Have Their Own Sandboxes – BigGo Finance (7 October 2026, 19:08 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind