Testing Center response evaluation
Key Takeaway
Response Evaluation runs by default in every test. It uses an LLM judge to compare what the agent actually says with an expected response. A score of 0 means a generic reply that did not address the user's intent.
Testing Center runs test utterances against an Agentforce agent and scores the results with evaluations. Some evaluations judge what the agent says. Others judge the choices the agent makes. Repeated test runs show quickly whether a change to the agent hurts outcomes.
The three default evaluations
Every test runs three default evaluations: Response Evaluation, subagent assertion and action assertion. Default evaluations are always tested. Response Evaluation judges the agent's words, but the two assertions judge its choices.
| What it checks | Score | |
|---|---|---|
| Response Evaluation | What the agent actually says, compared with an expected response | 0 to 5 |
| subagent assertion | The agent picked the correct subagent for the utterance | 1 (pass) or 0 (fail) |
| action assertion | The agent picked the right actions and included all expected ones | 1 (pass) or 0 (fail) |
The evaluations are independent, so one test case can pass some evaluations and fail others.
A user types "Book a table for four at 7 PM tonight." The agent replies, "Your table for four is booked for 7 PM tonight." The reply reads like the expected response, so it does well in Response Evaluation. The agent never ran the expected booking action, though, so the action assertion scores 0. The same test case passes one evaluation and fails another.