Anthropic Workshop: Build Agents That Run for Hours

youtube.com4 saved

Written by AI

TLDR
Ash Prabaker and Andrew Wilson argue self-evaluation is a trap, structured handoffs beat context compaction, and traces are the primary debugging loop for agent harnesses.
Filed note
Graph engineering for harnesses that stay up for hours. Came through a friend.

videoThe source

Moments

  1. Tuning a standalone critic to be harsh is tractable. Tuning a builder to be self-critical isn't.

  2. Before any code, the generator and evaluator negotiate what done means in files on disk. The evaluator grades against that contract.

  3. The main debugging loop was reading traces to find where the agent's judgment diverged from theirs, then fixing the prompt.

  4. The five takeaways: self-evaluation is a trap, compaction isn't coherence, use structured handoffs, write taste down as a rubric, read the traces.

Media saved