This week OpenAI published "Inside the Codex sandbox" and the Claude Code team published how they verify agent-written code.
What the Codex post covers
OS-level isolation (Seatbelt, bubblewrap + seccomp), a domain-filtering proxy, and permission profiles you can test with codex sandbox.
What the Claude Code post covers
End-to-end, property and differential tests, a separate grader, and verifiers kept out of reach of the changes they judge.
The gap between them
A sandbox limits what the agent can do. Verification checks the output. Neither gives you a record of what the agent did during the session to compare with its summary.
What I do about it
(Your own workflow here. If you mention RayTrace, add: disclosure, I help the founder.)
Top comments (0)