Sponsor talks on trustworthy evals and verification, vLLM/inference internals, microVM sandboxes, MCP, and autoresearch.
Accessible with the Expo Explorer pass and above.
This talk shows the alternative: every agent session in its own microVM, with its own kernel and a hard boundary to the host. You decide what lives inside that boundary: filesystem, network, the tools it's allowed to call. The sandbox runs Claude Code, Cursor, Codex, or whatever else you're driving. You'll see an agent live in full YOLO mode inside a sandbox, a real attempt to escape, and the boundary that holds up.