AI ROI: what your agents actually deliver

Worked out for 7 days, 30 days, 90 days or 1 year.

• Tasks (agent conversations), total tokens and active time.
• Success rate: tasks approved by Verify or code review, plus those that passed on the first check or review.
• Tokens per successful task, median time per task and messages after the first prompt.
• Tokens saved by smart tools and RTK.

Cards separate what works from what doesn’t.

• By model: success, first-time pass, tokens per successful task and time for each model.
• Project map impact: conversations that used the map against those that didn’t.
• Commits that needed fixing: agent commits against human commits; fix or revert commits on the same files within 14 days count.
• Reopened PRs: agent PRs against the rest.

The prompt cache card shows the share of input read from cache, the tokens it saved and why the cache was rewritten (waits over 5 minutes, or model and context changes), and lists large conversations that used it least. You also see code review findings, tasks finished and sent back on the task line, the CI failure rate on agent branches and risky migrations found by DB Guard.