Loops in the loop
This clock renders one real engineering mission. A human steered a Claude Code maestro through three days of work, weighing in at the moments that mattered and letting the rest run on its own.
The maestro dispatched 71 subagents across 19 workstreams in about 15 active hours spread over those three days: planning, building, reviewing and shipping, with the human stepping in 36 times to steer or approve.
The claim behind the picture: reliable agent-driven software comes from loops nested inside loops — the human and the maestro, the maestro's own work loops, an execute → judge → fix → re-judge cycle inside every workstream, and every agent's own loop of tool calls.
Position around the clock is mission progress — one notch per completed iteration, 146 in all — never wall-clock time.
On the right, each of the 19 workstreams is redrawn as its own complete loop, from its first dispatch to its last agent finishing. Its mini-clock locator shows where that workstream sits on the big clock.
How to read it
One wedge of the clock, rim at top (2×)
One workstream, redrawn as its own loop (2×)
Key
Four loop levels
Role (coil colour)
Model tier (coil stroke)
Verdicts (shape + colour)
Maestro marks
Bands and multiples
Scale
Workstream index
| Workstream | Type | Agents | Tool calls | Rejection loops | PRs | Loop story |
|---|