A convincing demo proves that an agent can work once. An evaluation proves where it works repeatedly and where it must stop.
Build scenarios from real work
Collect normal cases, edge cases, incomplete inputs, conflicting data and known historical mistakes. Remove sensitive information, but preserve the decisions that made each case difficult.
Score the whole behaviour
Measure whether the task was completed, the correct tools were chosen, arguments were valid, policy was followed, evidence was used and the final output was appropriate. A polished answer can still hide a dangerous tool call.
Test failure and recovery
Disconnect a dependency, return stale data, time out a tool and deny permission. The agent should fail visibly, preserve state and route the case correctly.
Set release gates
Define which failures block launch and which are acceptable with review. Run the same suite when prompts, models, tools or policies change.
Monitor production separately
Offline evaluations cannot predict every live input. Capture traces, feedback, corrections and exceptions, then turn production failures into new test cases.
The evaluation suite is part of the product. Without it, every improvement is also an uncontrolled experiment.