Agents Building Agents
summary
Claims a systems workflow: a coding agent (Claude Code) edits the product agent, runs self-healing evals against a golden dataset and scorers, and opens PRs or escalates for human input. An experimental loop called AutoAgent takes a bare agent from 18% to 83% pass rate in ~10 iterations, yields +10% on a human-optimized production benchmark, and raises another production agent from 67% to 86% in ~10 iterations. Harness engineering, tracing, clustering, and regression tests enable safe rollbacks.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:26AI agent challenges
- 3:31Golden datasets for agents
- 8:27AutoAgent improvements
- 13:07Optimization loop
- 19:04User feedback for agents
- 21:51User traces analysis
- 23:38Failure mode clustering
- 25:25Agent fix triage
- 27:58Agent reliability engineering
- 29:44Summary & Q&A
related talks
33mAgent Experience
45mAI agents in practice - beyond the prototype
30mArchitecting Integrity: Responsible Software Evolution in the AI Era
26mAI Agents as Your Co-Developer
40mHow To Tame Your Agents?
39mAgents for the Sake of Happiness
30mAI Agents Unchained
16mCoding Reinvented How AI Agents Are Shaping the Future of Software Development