Design Durable Long-Running Agent Queries
In 2026 design rounds, candidates have reported this prompt from July through October, at onsites, with the LLM and search APIs treated as black boxes you can call but can't change. Versions of it ask for the same things in different words: durable jobs, status and results the client can fetch later, resource limits, and clear behavior when a step fails, gets retried or gets cancelled.
Picture what happens today when a query takes twenty minutes. The agent plans, searches, reads a dozen pages, calls the model again, and starts writing. Somewhere in minute fourteen a deploy rolls the server. Every bit of that work lived in one process, so it's gone. The user sees a dead spinner, and when they retry, the query starts over from scratch and every model call gets paid for a second time. The design fixes that by moving the agent's progress out of memory and into a record that any machine can pick up. Most of the hour goes on what that record holds, when it gets written, and what happens around each external call. The scale is modest compared to search itself. The difficulty is correctness under failure.
PerplexityPremium lesson · System design round
The Perplexity design, worked end to end.
- Requirements and scopeSystem design · Signature problem
- The API and data model
- The design, drawn out1 diagram
- 6 deep divesWith the tradeoffs
$49 a year · all 107 companies · instant access
20,000+ candidate reports · guide updated
Perplexity interview guideKnow what Perplexity asks before you walk in.
24 coding · 5 system design · 22 behavioral questions
Get their questions first.
InterviewLegend is independent and not affiliated with or endorsed by any company it covers.