# Why AgentClash

Agent evaluation should test what an agent does across a complete trajectory, not only what one model call says.

Source: https://www.agentclash.dev/why
Markdown export: https://www.agentclash.dev/md/why

## The evaluation gap

Prompt tests miss tool choice, recovery, side effects, timing, cost, and multi-turn execution. AgentClash evaluates the complete run with evidence that teams can inspect and replay.


## The operating principle

- Same task and budget for every candidate.
- Fresh isolated environment for every run.
- Deterministic and model-based scoring with visible evidence.
- Failures become regression coverage instead of disappearing into dashboards.