Four eras in Lee's account

  1. Copy code from ChatGPT into an editor. 2:34
  2. Use terminal agents that edit files and run shell commands. 3:03
  3. Run coding agents through dedicated apps, sometimes in parallel or in the cloud. Reviewing every command still makes the human a bottleneck. 3:22
  4. Use ambient agents with their own computers, memory, and reusable skills that keep working while the user is away. 3:44

Interaction changes

  • Steering replaces a strict prompt-wait cycle. A user can ask the agent to check a thread, then immediately add that it should happen Thursday; the model and harness must interpret the second message as a correction to the same task. 4:08 4:29 4:53
  • Auto review uses a separate model to check commands. Lee argues that this can outperform humans fatigued by repeated approval prompts, but gives no benchmark or numerical result in this passage. 4:53 5:20
  • The product becomes a familiar texting interface for delegating real work, rather than only a question-answering interface. 5:38 6:05