Google DeepMind’s managed agents put cloud sandboxes and credential controls in one API

Google DeepMind engineer Jack Wotherspoon demonstrated managed cloud agents built around reusable sandboxes, multi-agent hand-offs and a proxy that keeps real credentials out of the sandbox. The pitch is practical: developers can compose agents without first building the infrastructure and security scaffolding themselves, though the demo gave no prices or performance figures.

Google DeepMind Watch analysis

What happened

In a podcast interview published on 26 October, Wotherspoon walked through managed agents using Google’s Interactions API. He described two stateful building blocks: an interaction that holds the conversation, and an environment that stores files and other work products. A second agent can reuse that environment with a fresh conversation, letting specialised agents pass work through files rather than inheriting one another’s context.

His demonstration had one agent gather NVIDIA filings, another analyse the saved data, and a third turn the results into a report. Wotherspoon also showed scheduled triggers and said the API’s credential proxy gives the sandbox a placeholder token, injecting the real credential only when an outbound request matches its destination.

Why it matters

Cloud sandboxes and managed credentials could make it easier to put agents inside applications without handing them direct access to secrets or asking every developer to build a container, harness and firewall from scratch. Reusing an environment while keeping each agent’s conversation separate also offers a straightforward way to split work into narrower jobs.

That is a developer-tooling demonstration, not proof that the approach is secure for every credential type or reliable at production scale. Wotherspoon gave no pricing, latency figures or failure-handling details.

Our read

The interesting bit is not that several agents can take turns. It is the attempt to package the awkward plumbing around them: persistent workspaces, scheduled runs and credentials that stay outside the sandbox. That could lower the barrier to building agent-powered products. Before anyone hands over the keys, though, the proxy’s actual coverage and the bill for all those API calls deserve a clear answer.

What to watch

• Whether Google publishes pricing and limits for managed-agent runs and triggers.
• Which credential types the proxy supports, and what protections apply to each.
• Independent evidence on reliability, failure recovery and the security of shared environments.

Discussion spark: Would you trust a managed-agent platform to keep credentials safe enough for production, or should teams run that security plumbing themselves?

not affiliated with, endorsed by, or operated by Google or Google DeepMind