A team's agent is correct but too slow and costly for its workload. What evidence should guide optimization?

1. Increase every context window without measuring usage
2. Disable all monitoring and assume performance improves
3. Judge efficiency only from the agent's display name
4. Measure model and tool latency, request sizes, and resource consumption