Notes on building agents.
Field notes from designing and operating production LLM agent systems — architecture, trade-offs, and the unglamorous parts (evals, failure modes, reliability).
Field notes from designing and operating production LLM agent systems — architecture, trade-offs, and the unglamorous parts (evals, failure modes, reliability).
Why a panel of specialist agents beats one model spread thin, how routing and structured output hold it together, and the eval mistake that made a dumb heuristic look smarter than a real LLM.