Five AI agents at work with a human in the loop, taking Kestrel Medical, a generated test company, from business problem to KPI dashboard. Each output appears in the middle as it's made, and a line traces every move of the work between agents. Gold lines are handovers after the Human approves a stage.
Final total 2h 46m. Claude also spent 36m improving the agents and their guides during the run, so 3h 22m of working time in all. Estimates: each step is matched to an agent by the files it touched.
Times are the recorded save time of each output file in the Kestrel Medical folder, the approval snapshots in History/, and the git commit times. A file that was saved again later shows its last save, so a few first drafts started earlier than shown.
| Source | Used for |
|---|---|
| File save times | Each output's save time in the project folder (or its first git commit, where that is earlier) |
| History snapshots | Stage approvals: the moment the approved version was saved |
| Session logs | Agent run time: only the time Claude was working inside a reply, with any single wait over 5 minutes cut to 5. Reading and review time between replies is left out. Each step goes to the agent whose files it touched; agent instructions, guides and personas count as improving the agents |
| Timeline | Steps are spaced by agent run time, not clock time, so long breaks and review time take no space. Each step also shows when its output was saved |
| Interactions | Each message Human typed to the agents in the sessions for this project, counted up to each step. Commands like /wrap and system notes are left out |
| Tokens | Every model call in the run's session log, up to each step: new input, cached re-reads and output. The main chat counts as Ovy and each subagent as its agent. The test stakeholder and helpers are left out |