Suppose an AI writes property listings from photographs, floor plans, and facts about each home. Is a new prompt actually producing better descriptions? Evaluation means assessing outputs against criteria chosen for the task. It can compare prompts, models, or versions of the same model, rather than relying on a few appealing responses. section 7.7 section 7.8
A useful evaluation starts by making success specific:
- Choose criteria. A property description might need to be accurate, tailored, positive, within a length limit, and clear about required facts such as flood-zone status. section 7.8
- Gather sample inputs resembling the intended task, then test candidate prompts on them. section 7.8 section 7.12
- Score outputs against the criteria. A metric is a score for a particular criterion; several metrics may be needed to understand performance. section 7.8
This process makes prompt engineering an experiment: change a prompt, compare its responses with those from the previous version, and look for improvements on the chosen examples. A change to a system prompt or model can also alter responses in unforeseen ways, so one attractive result is not enough to establish an overall improvement. section 7.7 section 7.12
When the AI acts instead of only answering
An AI agent may browse websites, use tools, and complete a multi-step task. The GAIA benchmark tests capabilities including reasoning, coding, and tool use through questions with unambiguously correct answers. That makes answers easier to score, but does not show how an agent reached them. AssistantBench focuses on tasks involving multiple websites; reported failure cases include relying too heavily on search results, inventing an answer, or failing to answer. section 8.16
For agents, a correct final answer may not capture the whole risk. Evaluation may also need to follow behavior over time and stress-test actions, including planning failures, memory drift, and misuse of tools. These methods matter when deciding whether an agent is reliable enough to deploy, though robust safeguards remain an open challenge. section 8.16 section 8.26 section 10.3
