2 min lesson
The last stake is adversarial
Put yourself in this case: "On the agent backend, why is 'just retry failed requests three times' an incomplete answer at Cursor's scale?" Give the clearest next step.
Step 1 of 2
The last stake is adversarial. A developer tool with repo access is a high-value target, so the security surface is large and someone is always probing it. Auth design here is defensive engineering: it has to resist a motivated attacker who benefits from every mistake.
Webhook ingestion and agent workloads don't arrive smoothly. A provider replays a backlog, a popular repo triggers a fan-out, a region wakes up - and you get a spike that's ten times the baseline in seconds. Designs that assume average load fail here. The interview-grade answer leads with how the system sheds, queues or degrades under the burst, then talks about the happy path second.
Don't reduce reliability to "add more nines." At 1M+ DAU on an LLM backend, the senior move is naming the cost of those nines explicitly (the warm capacity, the retry budget, the multi-region spend) and showing you'd pick an availability target on purpose. Treating dollars as someone else's problem signals you haven't owned a backend with real per-request economics.
“On the agent backend a retry is a spend decision as much as a reliability one. I cap it with a retry budget, use backoff with jitter so failures don't synchronize into a thundering herd, and make the call idempotent so retrying is safe. Three blind retries on an expensive long-running request just triple spend on the calls already struggling.”