Record and Replay

Record and Replay is the OpenAI computer-use mechanism described in 171: 【AI季报 26Q2】从 coding 到 RSI，强者愈强的未来？. The source says the system records a human completing a computer task, turns the workflow into a skill, and later lets the AI replay or adapt that process automatically.

The episode compares this to robot teleoperation: a human demonstrates an action path, and the system converts the demonstration into a repeatable capability. Its significance is that Computer Use Agent progress may come not only from better visual reasoning, but also from workflow capture, skill libraries, and an Agent Harness that can replay actions safely.

Key Claims

• Recording real human workflows can transfer tacit GUI operation into an agent-readable routine.
• Repeatability depends on interface stability, task boundaries, credential handling, latency, and recovery from changed screens.
• Privacy and permission rules matter because recorded workflows may expose accounts, documents, and business processes.
• Record-and-replay systems still need Human Judgment Under AI and verification when tasks affect money, code, customers, or regulated data.
• The route is more structured than free-form GUI clicking, but less flexible than a fully general reliable computer-use agent.

Connections

• OpenAI and Codex — company and agent context in the source.
• Computer Use Agent — broader agent category.
• Agent Harness, Agent Permission Boundaries, and AI Coding Verification — infrastructure and safety context.
• Claude Tag — parallel Q2 route for turning agents into repeatable workplace participants.