Most teams open agent traces when a customer complains. That's backwards. Technical failures are easy to catch. A tool call errors → you see it → you fix it. What actually kills agents in production is behaviour. The agent takes a worse trajectory. Quality drifts. It hits a tool-call limit and quietly degrades instead of stopping. No error. Just worse work. The fix isn't better prompts. Treat agent runs like a data pipeline: → Structured traces, captured by default → Auto-tagging that makes each pass more useful → Deterministic + non-deterministic logs together → Evals continuously running on real production runs → Human feedback steering AI evaluators The goal isn't just observability. It's making every run teach the system something. There's another problem almost nobody solves: The person best qualified to judge whether an agent did good work is often a domain expert who can't read a trace. Today, their feedback ends up in a Slack thread. A week later, it's gone. That feedback needs to become structured data the system can actually learn from. I've spent a decade in data engineering, then started building agentic systems where non-technical users found weak points within days. Agents are a data problem wearing an AI costume. Building the tooling for this at @neatlogs.