quiz · 4 min Observability & Debugging Check Why does Lesson 42 say print debugging is insufficient for agent loops?Agent runs are non-deterministic, so the exact failing run cannot be re-created by re-running the codeYou can reconstruct the failing run from print output after the factRe-running with the same input reliably reproduces the failureFree-form log strings capture all the information needed for debuggingWhat is a trajectory, as defined in Lesson 42?A single event capturing the agent's final answer and the total run durationA real-time alert that fires when a step produces an errorThe run id minted at the start of a run and propagated to every eventThe ordered sequence of events for a single run -- a complete, queryable record of what the agent saw and decidedWhat does the Tracer.step() method do in the recorder built in Lesson 43?It replays stored observations through the policy and checks for action regressionsIt invokes the policy to decide the next action from the current observationIt appends a structured record of the observation and action to the events list and returns the actionIt formats the observation and action as a log string and prints it to the consoleAccording to Lesson 44, what is the correct first step when a stored trajectory is available for a production failure?Re-run the agent with the same input to reproduce the failure in a live environmentLoad the stored trajectory, replay it by feeding each recorded observation through the current policy, and compare actionsAdd more structured event fields to the recorder so the next run captures more detailIdentify the fault layer -- Perception, Policy, or Tool execution -- before touching codeAfter isolating the bad step in a traced failure, what does Lesson 44 say you should do next?Immediately patch the policy layer since that is where agent decisions are madeRe-run the agent live to confirm the fix produces the correct final answerAdd an alert threshold on the step error rate to catch the same failure earlierExtract the (observation, expected action) pair from the bad step and add it to the evaluation suite as a regression caseSubmit Mark complete ← Debugging From Traces Shipping an Agent →