· Perseval · Agent evaluation · 2 min read
Why Agent Trajectories Matter More Than the Final Answer
Outcome-only scoring misses the decisions, tool calls, and recoveries that make agent behavior reliable or fragile.
Two agent runs can produce the same final answer through very different trajectories. One may plan cleanly, use the right tools, and verify its result. The other may loop, leak context between roles, call the wrong tool, and recover by chance.
Outcome-only evaluation treats them as equivalent. Operationally, they are not.
What a trajectory exposes
A trace can show:
- which agent or role made each decision;
- how tool arguments changed over time;
- where execution diverged from a healthy baseline;
- whether a verifier caught an error;
- whether retries recovered or merely concealed a fault;
- which evidence supports a failure diagnosis.
That context is especially important for multi-agent systems. Flattening planner, browser, tool, model, and verifier activity into one list removes the topology needed to understand ownership and causality.
Perseval preserves the hierarchy, events, links, attributes, and payload references needed to investigate the whole run while keeping large traces navigable.
The sanitized handoff scenario provides a concrete example. Planner, browser, and verifier spans share one orchestrator parent, while the verifier links to browser evidence. Tree shows ownership, Timeline shows overlap, and the SpanLink shows consumption. No single flat sequence can express all three relationships honestly.
The recovered-optional scenario adds another topology lesson: an optional browser lookup errors, then a retriever supplies the required fallback. The browser status stays red because that fact happened. The completed root and successful fallback explain why the same red span need not become an actionable finding.