What is baseline measurement?

Example: You have an AI agent that summarizes customer support tickets and routes them to the correct department. Response quality feels inconsistent, so you decide to optimize the prompt. You rewrite the system instructions, add few-shot examples, and switch from a general model to a specialized one. The summaries look better to you. You declare the optimization a success and move on. Three weeks later, routing accuracy has dropped. Customers are complaining about misrouted tickets. What happened? You have no idea, because you never measured anything before you started changing things. You do not know what the baseline routing accuracy was. You do not know what the baseline summary quality was. You do not know whether the changes you made improved some dimensions while degrading others. You optimized without a benchmark, so you cannot distinguish between actual improvement and the feeling of improvement. Now rewind. Before touching anything, you run the agent on 200 historical tickets and score the outputs: routing accuracy is 74%, summary completeness is 68%, average latency is 2.3 seconds. You record these numbers. Then you make your changes. You run the same 200 tickets through the new version: routing accuracy is 81%, summary completeness is 79%, latency is 2.8 seconds. Now you know exactly what improved, by how much, and what tradeoff you introduced. The optimization sprint in Optimization sprints gave you dedicated time for improvement. The benchmark gives you proof that the time was well spent.

This concept is part of Agent Optimization in the How to Think curriculum.