Debugging scenarioHard~15 minAgents & Workflows / 17 of 18

Interview question

An agent keeps repeating tools without making progress. How would you diagnose it?

Uses traces to separate unclear goals, bad tool interfaces, missing state, contradictory results, poor termination, and retry amplification.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

I stop or isolate the run, then inspect the full trajectory: goal, instructions, eligible tools, calls and arguments, results, plan changes, progress markers, and termination checks. I look for identical calls, equivalent calls with rewritten arguments, errors returned as success, missing observations, and state that was never persisted. The immediate safeguard is a repetition and no-progress limit.

Full answer below

Practice focus

Agent debugging, trajectory analysis, loop detection, tool contracts, and durable correction.

Answer first

Answer path

Core answer first, then checkpoints toward a stronger answer.

Start
Core answer
Start here. This is the direct spoken answer to practice first.
Overview
Repeated calls are a symptom; the cause may be the agent, the tool, the state model, or the task contract.
I stop or isolate the run, then inspect the full trajectory: goal, instructions, eligible tools, calls and arguments, results, plan changes, progress markers, and termination checks. I look for identical calls, equivalent calls with rewritten arguments, errors returned as success, missing observations, and state that was never persisted. The immediate safeguard is a repetition and no-progress limit.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
Common causes include overlapping tool descriptions, a tool result that omits the field needed for the next decision, stale context, an impossible completion criterion, blind retry, or a plan that forgets completed work. I reproduce the case with captured tool responses and compare it with successful trajectories. The fix targets the boundary: improve schemas, return typed errors, persist state, narrow tools, or make completion deterministic.
Finish
Advanced answer
Optional deeper angle for production judgment and nuance.
I check whether concurrency, eventual consistency, or an ambiguous side effect makes the agent observe old state and repeat a dangerous action. Tool receipts and idempotency prevent further harm while reconciliation resolves the real outcome. Evaluation adds the failure as a trajectory case and measures unnecessary calls and time-to-stop, not only eventual completion. Production detection clusters repeated tool and argument patterns by release version without retaining more sensitive content than necessary.

Review signals

Use these to catch weak phrasing, shallow reasoning, and risky claims.

Common mistakesCandidateCommon mistakesMistakes that weaken the answer or signal shallow understanding.Open details

Details

Increasing the maximum-turn limit.
Rewriting the prompt without inspecting tool results.
Assuming every repeated call is harmless.

Red flagsCandidateRed flagsSignals that an interviewer may treat as risky.Open details

Details

Tool errors are returned as ordinary success text.
No progress marker exists.
Repeated side effects lack idempotency.

Follow-up prompts

Quick interviewer prompts inside this drill.

Follow-up questionsCandidateFollow-up questionsLikely next questions after the main answer.Open details

Details

Treat these as the next interviewer turn. Answer one out loud, then save the gap in your practice notes.

How would you define progress?
Which trace fields matter most?
What if the external system is eventually consistent?

Continue the section

Next in Agents & Workflows: Design a bounded agent

Step 18 of 18

Design a bounded agentic workflow for a production application.

Practice route

Read core answer Review checklist Follow-up prompts Track progress

Finished this drill?

Practice freely. Sign in only when you want to save completion, confidence, notes, or bookmarks.

New to Aporeon? Create account

Drill details

Human editorial direction

Aporeon is shaped by Aleksandar Tomovski, a software developer with experience on both sides of technical interviews. Content is reviewed for accuracy, natural spoken delivery, useful depth, and honest trade-offs.

About the creator Editorial standards

Was this drill useful?