Most agent demos run for thirty seconds. Production agents run for hours, and almost everything that breaks at that timescale has nothing to do with the model. Drawn from building an agentic data-QA engine running at 20B+ tokens per week in production. Why long-running agents fail: context degradation, silent tool failure, unrecoverable state, cost drift, and why demos never catch them. How harness design answers each one: compaction and memory, per-role model routing, sandboxed execution, checkpointing and recovery. Where to start: what to build first moving an agent from demo to production, and what to instrument on day one.
AI/ML Engineer at Zyte, technical lead for LLMNL, an AI-powered data-QA engine running agentic workloads at 20B+ tokens per week. Previously led Promtior's Augmented Behaviors Lab (applied R&D in agent memory, voice and multimodal). Teaches Applied AI at ORT University, Montevideo.
