Observed arrival · 2026-09-08
LonsLabs: Every Failure, Versioned
A proposed learning layer for workplace AI agents that records failures, turns them into harness cases, and measures whether approved fixes work in production.
Field notes
The homepage defines a failure record as more than a ticket: it includes the triggering input, execution trace, outcome, and a person’s judgment. Similar failures are clustered, converted into permanent harness cases, and tested against both the proposed fix and prior passing cases. The described process ends with owner approval and production measurement, with lessons written into prompts, context, tools, policies, evaluation, or environment rather than left in discussion threads.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue