Discover conversational-agent failures

Find specific ways the agent fails its job, distinguish them from bad test execution, and turn the findings into a useful next test. This is exploratory analysis; an agent's judgments are proposals, not human labels.

Scope and sample

Confirm the product question, organization/workspace, agent and time window or local source file. Reuse available evidence before generating more calls. Use coval --agent agent manifest to discover reads. Simulated conversations and uploaded production conversations are different collections:

coval --agent simulated-conversations list --run-id <run-id> --page-size 100
coval --agent simulated-conversations get <simulation-id>
coval --agent simulated-conversations metrics <simulation-id>
coval --agent uploaded-conversations get <conversation-id>