Architecture scenarioHard~15 minAgents & Workflows / 15 of 18

Interview question

How would you recover a long-running agent after a process failure?

Checkpoints authoritative state, reconciles ambiguous tool outcomes, preserves approvals, and resumes without repeating completed work.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

I persist the run goal, current state, completed steps, tool requests and receipts, approvals, budgets, and terminal status at durable boundaries. On restart, the runner loads that state and continues only unfinished work. Tools use idempotency keys or reconciliation queries so an ambiguous timeout does not repeat a side effect.

Full answer below

Practice focus

Durable execution, checkpointing, idempotency, reconciliation, approval persistence, and resume semantics.

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
Recovery means resuming from known state, not replaying the entire conversation and hoping the model repeats itself.
I persist the run goal, current state, completed steps, tool requests and receipts, approvals, budgets, and terminal status at durable boundaries. On restart, the runner loads that state and continues only unfinished work. Tools use idempotency keys or reconciliation queries so an ambiguous timeout does not repeat a side effect.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
Checkpointing occurs before and after consequential operations and whenever the run pauses for a person or external event. In-flight work has explicit unknown outcomes that must be resolved before replanning. Model-generated summaries help rebuild context but do not replace the authoritative event and state record. Resume verifies that tool schemas, permissions, model policy, and approvals are still valid.
Finish
Advanced answer
Optional deeper angle for production judgment and nuance.
The recovery model handles code deployment, model change, tool retirement, revoked identity, expired approval, and changed resource state. A migration layer upgrades durable run state or deliberately terminates incompatible runs with a clear recovery path. Budgets do not reset automatically after restart. Traces join pre-failure and resumed execution under one run identity, while replay tools can simulate decisions without re-executing side effects. Recovery testing kills workers at every critical boundary.

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

Replaying the full agent loop from the beginning.
Treating timed-out actions as failed without reconciliation.
Resetting step and cost limits after restart.

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

Details

The only state is conversation text.
Approvals disappear or execute twice after restart.
Old runs invoke changed tools without compatibility checks.

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.

Where would you checkpoint?
How do you resolve an unknown tool outcome?
What can invalidate a resumed run?

Continue the section

Next in Agents & Workflows: Evaluate an agent

Step 16 of 18

How would you evaluate an AI agent end to end?

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?