Production patterns for Forward Deployed Engineers.
Why tool calls fail, how to make retries safe with idempotency keys, and where to put them: client, server or both.
How JSON mode and schema-constrained output differ, and how to make LLM tool calls parse reliably in production.
How agent hallucinations differ from model errors, how to measure them, and the checks that catch them before customers do.
How the loop terminates on stop_reason, and why a coordinator delegating to isolated subagents beats a flat swarm.
Bigger windows feel like free memory, but accuracy decays as input grows, even on trivial tasks. The evidence across 18 models, plus what to do about it.