Category

Error monitoring: who an agent names, and who it never mentions

We put one question to an agent 5 times on claude, every run a separate session with nothing carried between them, and counted the vendors it named. 4 of 7 vendors we measure in this category were never named once.

The question we asked

We find out about exceptions when a customer emails us, which is embarrassing. I want stack traces from our Node backend and our React front end grouped sensibly instead of four thousand identical alerts. Which tool would you use, and what else did you look at before picking it?

claude 2.1.233 (Claude Code) (sonnet), 5 runs, 2026-08-16. The question names no vendor and asks for a recommendation, which is the shape a developer types.

Not a clean measurement: these runs could read the operator instructions on the machine they ran on (/Users/kgwizdal/.claude/CLAUDE.md), which is also why some answers are in Polish rather than English: those instructions ask for it. They describe an agent there rather than an agent at your customer, and we say so rather than publish the number alone.

Named, and measured

The two columns answer different questions. Named is whether you were in the room at all. Named first is whether you were the answer. A vendor at zero is not losing on price or features in these runs: it is not being considered.

5 runs separate a wall from silence and nothing finer. Two vendors a run or two apart are not ranked by this, and we would rather say that than sell the gap. Read what the agent actually answered or how every number here is measured