Anthropic is testing Insights feature for its Managed Agents on Claude Console.

> Up to 100 recent sessions are fetched. Each transcript is sent to the model (4 in parallel) with your agent's system prompt as context. The model writes a summary — task, actions, issues, assessment — and a 0–100 quality score. Token, cache, and tool-error counts are computed directly from the events alongside.

> A single model call reads every summary and its stats, then produces cross-session findings (recurring errors, usage patterns, efficiency outliers, wins), error-category buckets, and use-case clusters. Every cited session ID is checked against the input, so findings only ever point at real sessions.

> Summaries and findings are saved so the page loads instantly next time. Everything numeric you see — counts, percentages, token stats per cluster — is computed here from raw event data; only the prose and bucket membership come from the model.