Your AI Agent Is Hiding Its Mistakes From You - Ben Holmes, Warp

By ai.engineer

Categories: AI, Tools

Summary

AI agents systematically hide their failures by not surfacing mistakes to their operators. Ben Holmes reveals how to build self-improving agent loops that automatically review past work, score quality against defined rubrics, and suggest configuration improvements—turning hidden failures into optimization signals.

Key Takeaways

  1. Centralize all agent conversations and metadata (tool calls, file system calls, MCPs, code diffs) in a single accessible location to enable systematic introspection and failure analysis across your entire team.
  2. Implement a four-stage improvement loop: (1) collect agent conversations, (2) score quality using LLM-as-judge against custom rubrics (efficiency, code quality, test coverage), (3) identify failure patterns, (4) auto-suggest improvements via cron-triggered agents.
  3. Use scoring agents as judges to evaluate agent performance against defined rubrics rather than relying on manual inspection; this scales failure detection and creates quantifiable signals for optimization.
  4. Factory engineering (meta-engineering around agents) requires deliberate optimization loops—without systematic review, agents continue repeating the same mistakes because failures remain invisible to the system.
  5. Software factories running on shared cloud infrastructure (Docker sandboxes, isolated compute instances) automatically centralize agent traces, making self-improvement loops easier to implement than distributed local agent setups.

Related topics

Transcript Excerpt

I'm Ben Holmes and I'm a DevEx engineering lead at Warp and I'm going to be talking about building self-improving agents, agents that look through their past work and figure out how to improve their own setup. Uh so for a bit of context, I work on Warp Factories, which is open flexible infrastructure to help you all build software factories of your own and crucially ones that improve themselves as part of the stack. Uh I'm not going to be talking too much about software factories as a whole today, but I'm going to be talking about one crucial part of it, which is having agents introspect all of the work that they do inside of one of those systems and suggest improvements so that the agent performs better and better over time. Uh so I'm going to start by just presenting what you might think…

More from ai.engineer