Agent evaluation extends LLM evaluation from single answers to whole tasks. An agent plans, calls tools, reads data, and takes actions over several steps, so evaluators look at two levels. Outcome checks ask whether the task ended in the right state, such as the correct refund issued. Trajectory checks review the path: which tools the agent called, in what order, with which arguments, and whether it stayed within its permissions. Teams also track cost, run time, and how often results change across repeated runs.
Public benchmarks compare agents on shared tasks, but they rarely match a company’s own systems and records. Many teams therefore build a labeled set of tasks from their own work, often called a golden dataset, and rerun it under fixed conditions whenever the agent, the model, or the underlying data changes, so results stay comparable over time.