Testing & Reliability
Evals, debugging, tracing, why agents fail.
30 posts · showing 1–10
-
Your Research Agent Scored 100% on Citations. The Grader Only Checked That a Source Was Attached.
An eval that has never failed anything is not proof of a good agent. Before you trust a grader's score, feed it answers you know are wrong and see if it notices.
-
Your Research Agent Passed the Eval Once. Run Every Case Five Times and Grade the Worst Run.
A single green run tells you the agent can do the task, not that it will. Repeat each eval case, split the suite into solid, flaky and broken, and the flaky bucket becomes your real to-do list.
-
The Agent Passed the Test. Unfortunately, It Tested the Wrong Thing.
AI agents are getting very good at completing tasks—and very good at satisfying badly designed tests. The dangerous part isn't a failed test; it's a green test suite that gives you false confidence.
-
The Agent Got the Right Answer by the Wrong Route. Score the Trajectory, and Know the Three Cases Where the Answer Is Enough
A final-answer eval passes an agent that guessed, one that ran a write query mid-task, and one that took forty steps to do the job in four. Here is how to assert on the path without freezing a golden trace, and when you can safely skip it.
-
Retrieval Returned Zero Rows and the Agent Cited a Policy Doc Anyway
Empty retrieval results are the only tool failure that arrives as a success — status 200, results [] — and a model asked to answer from sources will invent one to fill the hole. Two boundary fixes: a loud typed empty, and citation IDs validated in code.
-
A Second Pass That Asks the Same Question Finds the Same Bugs. Change the Unit, Not the Temperature.
Raising the temperature on a review pass re-rolls the dice inside a frame that has already decided what is findable. Three properties determine what a pass can see, and a real second pass has to change at least two of them.
-
A 1% Flaky Tool Shouldn't Take You Down. Three Nested Retry Layers Will.
A tool that fails one call in a hundred is healthy. The retry logic wrapped around it is what turns a dependency's bad minute into a queue of hung user requests — and the load multiplier is at its worst exactly when the dependency is weakest.
-
A Two-Line Prompt Diff Is a Behavior Change. Here's the Regression Check It Needs.
A prompt edit has the blast radius of a code change and none of the safety rails — no compiler, no type error, no failing test. Six checks that turn a forty-second approval into a real regression gate.
-
Your Eval Ran Each Case Once. That's Why It Missed the Flake.
A forty-case suite can pass every case and still hide a bug that reaches users one week in five. The number missing from most agent evals isn't coverage — it's how many times you ran each case.
-
Green Evals, Broken Production: The Frozen Fixture Problem
An eval suite built on recorded tool responses stops testing your agent and starts testing a museum. Here are three cheap detectors for fixture rot — and the rule for when a case should call the live tool instead.