Part 1 of 4 · 1 min read
Overview
An agent chains many calls, so per-call failure compounds: ten calls at 99 percent each succeed about 90 percent of the time. Without these patterns a provider blip becomes a user-visible outage and retry storms make it worse. Interviewers ask "what happens when your model provider goes down?" and expect a layered answer.
About the example, and what is real
Real: the circuit breaker (three states), retry budget, token bucket, retry amplification arithmetic, outage simulation and bulkhead simulation, and the 19 test cases in code/mini-reliability. Assumed: provider fails 100 percent (fast) for 120 s and 2 percent otherwise; fallback provider 99 percent available and independent; 20 requests per second; breaker thresholds (20 calls, 50 percent, 10 s); retry backoff; tool B hangs 30 s; token-bucket burst 5. Not real: no real provider, network, concurrency or clock; an outage with slow responses, partial failures or correlated failures behaves differently.