The result matters. So does the path.
Action-level evaluation that exposes brittle, unsafe or evidence-poor behaviour hidden behind a successful final answer.
Discuss a similar programmeA successful-looking answer that conceals the agent’s decisions and actions.
The performance gap
A final answer cannot show whether an agent ignored evidence, took an unsafe action or succeeded through a brittle sequence. Those weaknesses matter when the same behaviour has to operate inside real software.
Our contribution
We evaluated the trajectory behind the result. Expert judgement connected tool calls, state transitions and recovery behaviour with task-specific quality expectations, making hidden weaknesses visible to the model team.
Value delivered
The programme delivered environment-ready tasks, trace annotations, action-level critiques and success evidence. The advantage was a richer reliability signal: what the agent achieved, where its behaviour failed and which failures should shape the next training contribution.
Platform context.
Bring the next
requirement.