OpenAI Codex launches Record & Replay
OpenAI's Codex 26.616 update introduces Record & Replay, a feature that automatically compiles manual user actions into reusable skill packages. This coincides with a new paper detailing the limits of automated trajectory mining for agent policy optimization.
Hot take: Direct demonstration via Record & Replay is currently far more viable for agent optimization than unsupervised trajectory mining, which fails to reliably transfer to new domains.
- –Record & Replay simplifies agent teaching by recording human actions on macOS and converting them into a structured SKILL.md format.
- –Trajectory mining can produce highly readable clusters (achieving up to 95% purity against human benchmarks) but lacks policy transferability.
- –The research paper highlights that orderless segment representation and weak offline reward models are major bottlenecks for automated skill generation.
- –The SKILL.md standard continues to grow as the default cross-platform format for packaging and steering AI agent behaviors.
DISCOVERED
2h ago
2026-06-19
PUBLISHED
2h ago
2026-06-19
RELEVANCE
AUTHOR
omarsar0