Why Your Agent Evaluations Will Fail You (and How to Fix Them Before Production)
Notes de l’épisode
Anthropic deprecated Sonnet 3.5. Some of Xelix's pipelines migrated smoothly. Others broke — and customers noticed within hours. What separated the two? Evaluation. Paul Solomon and James Price Farr have spent 5+ years building AI systems that process millions of invoices for enterprise customers. In this episode, they share the evaluation-first framework that now saves them every time a model changes, an orchestration layer fails, or an agent picks the wrong tool. Key takeaways: • Evaluation-first, not evaluation-after — Retrofitting evaluation on an agent already in production is painful. Build your eval pipeline before you build the agent. • Monitor tool calls, not just outputs — If the agent isn't selecting the right tools, nothing downstream will be correct. Tool-call monitoring is your leading indicator. • 3 tiers of automation — Not everything needs an agent. Rules-based → single LLM call → agentic system. Pick the simplest tier that solves the problem. • Extended thinking tames token explosion — After migrating to newer, more verbose models, enabling extended thinking (with a budget) moved reasoning out of expensive output tokens and brought costs back under control. • Human-in-the-loop by default — Start with human review on every output, then earn trust toward touchless automation as customers gain confidence. • Pragmatism wins — Use whatever technology works best for the problem. Not every feature needs an LLM. Recorded live at AWS Summit London.With Paul Solomon, Head of AI Engineering at Xelix ; With James Price Farr, AI Engineering Team Lead at Xelix