PLATFORM
Agent Evaluation.
Evaluate the action, not just the answer.
Agent evaluation should inspect outputs, tool decisions, trajectories, constraints and pre-execution state — not only text similarity.
Four evaluation layers
01
Output: Is the generated result correct, relevant and well-supported?
02
Tool use: Did the agent select an appropriate tool and parameters?
03
Trajectory: Did the sequence of steps create unnecessary risk or drift?
04
Action: Should the final external action actually be allowed?
Offline and runtime
Offline evaluations are useful before deployment. Runtime evaluation matters once agents operate on live context. Immortal Dynasty is centered on connecting those evaluation signals to a real decision boundary.
Use NexusEval as the engine
Start with bounded live checks exposed by NexusEval, then compose those outputs into your own policy and gating layer.
Explore NexusEval →