Anthropic engineer: over 90% of engineers build with self-improving agent loops

Roger_M_Taylor · x · 2026-09-19

An Anthropic engineer says over 90% of engineers at the company build with self-improving loops, heading to 100%. In a 40-minute breakdown he explains how the Claude team deploys agentic loops that run for days at a fraction of standard inference cost: Prompts → Harness → Self-Healing Loops → Execution Graphs → Autonomous Swarms. The point is building an architectural harness where parallel agents catch hallucinated code, isolate memory leaks and auto-recover from worker crashes — wiring feedback loops that test diffs and verify production state. The cited article 'Graph Engineering' argues most multi-step agents are serial queues that should be graphs.

Related event: Anthropic Engineer: 90% Use Self-Improving Loops for Agents(2 posts)→

Original post →

More from coding & agent

coding & agent channel →