A 311 resident-service assistant occasionally gets stuck repeatedly calling the same lookup tool with nearly identical arguments and never produces a final answer for the resident. What production issue is occurring, and what should the team check first?
Select an answer to reveal the explanation.
Short Explanation
Think of someone stuck re-reading the same page because they can't decide it actually answered the question — that's a reasoning loop. The agent isn't broken at the tool, it's stuck deciding when it has enough to answer. The fix starts with the trace, looking for where the stop condition should have fired and didn't.
Full Explanation
An agentic loop repeatedly calls a tool with near-identical arguments because its termination logic never recognizes the returned result as sufficient, so it keeps retrying instead of finalizing — a distinct production failure mode from an outright crash. It's diagnosed by reading the reasoning trace to find where the loop-exit condition should trigger and doesn't, then tightening the stopping criteria in the prompt or adding a hard iteration cap. A cascading failure implies the dependency itself is unavailable, but here the tool is answering each call fine; it's the agent that won't stop asking. Hallucination means inventing content absent from the source, not repeating the same real call over and over. Blaming the tool's own latency misreads a decision-logic failure as an infrastructure one, since a single successful call clearly isn't being recognized as complete. Caveat: reasoning loops can resemble simple retries after transient errors, so confirm the tool's responses are actually succeeding before concluding it's a reasoning defect. Operational check: cap iterations per request and alert when a request exceeds the expected tool-call count for its task type.