AG08:2026 — Cascading Failures

Slide 24 · Quiz — Questions 1 & 2

Test your understanding of cascading failures in multi-agent systems.

1. A retrieval agent returns 5% of its expected records due to an API timeout that returned a partial success. The analysis agent processes the truncated data and produces confident-looking output. No error signal is emitted at any stage. Which failure pattern does this most closely represent?

Pattern 4 — Resource exhaustion cascade, because the partial response caused extra processing load Pattern 1 — Silent error propagation, because the failure produced structurally valid output with no error signal at any hop Pattern 6 — Adversarial cascade, because an attacker caused the API to return partial data Pattern 3 — Runaway feedback loop, because the analysis agent re-read the retrieval agent's output

2. A vector database used by 4 retrieval agents has a 12-second outage. All 4 agents retry simultaneously every 2 seconds with no jitter. The database recovers at 12 seconds but is immediately overwhelmed by the simultaneous retry bursts and fails again. This repeats for 4 minutes. Which mitigation would most directly have prevented the extended outage?

MIT01 (output validation) — checking the agents' outputs would have caught the bad data MIT06 (rate limiting + human confirmation) — requiring human approval before retrying would have stopped the storm MIT04 (exponential backoff with jitter) — randomizing retry timing would have prevented simultaneous bursts from overwhelming the recovering database MIT07 (idempotent operations) — making retries idempotent would have prevented the duplicate requests from causing harm

← Back Questions 3–5 →