What's this about?
Wes McKinney (creator of pandas) posted "I think loops are bullshit," kicking off a debate about agent-harness design. Simon Willison's synthesis: explicit retry/verify loops were a temporary patch for models that couldn't stay on task — Fable and GPT-5.6 now "just do that out of the box."
What exactly is an "agent loop"?
The scaffolding that repeatedly re-prompts a model — try, check, retry — until a goal is met. Early agents needed the harness to be the persistence; the claim is that persistence has now moved into the models themselves, making elaborate loop scaffolding legacy code.
What's the community saying?
471 + 245 likes
- The harness-minimalism camp: every layer of scaffolding you add is a bet against the next model generation.
- The counter: loops are also about *auditability* and cost control, not just capability babysitting.
- It rhymes with yesterday's RLM-paper debate about whether generalization should live in the harness or the model.