Sandboxes Aren't Optional: 4 Ways to Isolate Coding Agents — Robert Brennan, OpenHands

Robert Brennan, co-founder and CEO of OpenHands, gave this 14-minute talk at the AI Engineer World's Fair 2026 in San Francisco. His case is that a coding agent should run in a sandbox, and he walks through four concrete ways to set one up.

After a short introduction to OpenHands at 0:00, Brennan traces how AI-driven development has changed (0:57): from code snippets copied out of ChatGPT, to agents on a developer's own machine, and then to agents in the cloud (2:27). At 3:32 he explains the risk of the local setup. An agent started on a laptop with the dangerously-skip-permissions flag can reach every file, secret and production credential its owner can. The principle of least privilege, introduced at 4:52, frames the rest of the talk.

The four setups follow in order. Docker containers on a local machine come first at 5:42. At 7:02 he covers Claude Code running inside GitHub Actions. A dedicated VM or Mac Mini is the third option (8:22), and a Kubernetes pod for each agent session is the fourth (9:47).

At 11:02 he compares the four side by side. According to the description, the comparison rates each option on local versus remote, scale, isolation, persistence and triggers. At 12:22 he explains why sandboxes do not solve everything, and the description names container escapes as one of the limits. The talk closes with a look at OpenHands Agent Canvas at 13:07 and a brief wrap-up at 13:57.

The comparison chapter at 11:02 is the fastest route to the trade-offs if you already know the options.