ERRORPROBE: automatically pinpointing which agent caused a multi-agent failure
burkov · x · 2026-10-18
The article introduces ERRORPROBE, a framework for debugging LLM-based multi-agent systems. In such systems, failures are hard to trace: errors may occur many steps before symptoms appear, and agents depend on each other's outputs. Existing diagnostics rely on costly expert annotation or a single LLM judge, neither of which can pinpoint the exact step where an error originated in long interaction traces. As multi-agent systems spread across software engineering and web navigation, ERRORPROBE automatically identifies both the failing agent and the specific step where the failure began.
More from coding & agent
- One founder's week on personal AI agents: taxes filed, Hamilton tickets, and an honest purchase scoreboard — altryne · 2026-10-18
- Shopify's Tobi Lütke: 'We are all verification engineers now' as agents become the third customer — schwentker · 2026-10-18
- Shopify CEO Tobi Lütke: "We are now all verification engineers" — schwentker · 2026-10-18
- Anthropic researcher on agent memory: "dreaming" aggregates memories from past sessions — schwentker · 2026-10-18
- Voice demo built with Jev draws "holy crap, that's the missing thing to get rid of keyboards" reaction — schwentker · 2026-10-18
- Agents reach llms.txt in only 32% of cases; 86% of those via site links — schwentker · 2026-10-18