OpenAI's managed Agents API raises the auditability question: what must stay inspectable?

Mediocre-Main-9870 · reddit · 2026-09-14

OpenAI's Agents API (public beta) has developers specify task, model, tools and environment while the provider hosts the harness; the underlying Codex harness is open source and inspectable. The poster argues the real production question is where the trust boundary sits: an inspectable harness doesn't reveal configuration, policy decisions, tool permissions, or run evidence.

He asks which artifacts teams should insist on retaining for incident reproducibility: resolved tool schemas and permissions, policy/approval decisions, input fingerprints, execution receipts, environment versions, or something else.

Original post →

More from coding & agent

coding & agent channel →