28 Feb 2026/3 min
The boring half of agents.
Everyone wants to talk about reasoning loops. Almost nobody wants to talk about retries, idempotency, and the queue that quietly carries the whole system.
Agentic systems look like LLM problems. They are mostly distributed-systems problems wearing an LLM costume.
The interesting failures aren't 'the model hallucinated'. They're 'the tool call timed out, we retried, and now we charged the customer twice'. Idempotency keys, dead-letter queues, and a boring observability layer fix more user-visible bugs than any prompt change.
If you're building agents and you haven't drawn the state machine on a whiteboard, you're not building agents — you're building incidents.