Agent-fleet intelligence: one human, N agents

I run a fair-sized fleet of coding agents most days, and the honest truth is that for a long time I couldn't have told you what it cost, or how much of what shipped was me and how much was them. The spend was a line on an invoice. The output was just "the work". For something I lean on this hard, that's not much to go on.

The obvious move, the one most tooling reaches for, is to point the microscope at the developer. Count their commits, clock their hours, rank them against each other. That's the wrong instrument aimed at the wrong question, and I'll cheerfully die on that hill. The research on developer metrics has been clear for years that measuring individuals like this gets gamed and quietly poisons a team, and when a person's job is increasingly to direct agents rather than to type, a keystroke count measures roughly nothing worth knowing.

So it points the other way. It sits on top of the routing receipts that OmnisRouter already emits, one per request, and rolls them up to the level that actually helps a lead: the team, or the department. What the agent spend was this week, and what it bought. Which models the work leans on. What a shipped change costs on average, once you count the tokens behind it. Where the agents earn their keep, and where they generate more review than they save.

# One OmnisRouter receipt per request; the dashboard aggregates them.
{ "task": "refactor-auth", "model": "sonnet", "tokens_in": 18240,
  "tokens_out": 3110, "cost_gbp": 0.061, "routed_from": "opus" }

team / aug:  4,182 requests   £248   71% routed down a tier   £0.12 per merged change

The hard part

The hard part isn't the plumbing. It's resisting the pull to turn this into a scoreboard. Every metric's an aggregate, the collection is opt-in, and every figure opens up to the receipts underneath it, because a productivity number you can't interrogate is just management by vibes. Hold that line and you've got a tool a team will actually trust. Lose it and you've built the surveillance product I set out not to build, and that slide is far easier to take than anyone admits.

MEASURE THE FLEET, NOT THE TYPIST.

It's in design, and further from a prototype than most of what's on this page. The near-term work is the data model over OmnisRouter's receipts, and working out the smallest set of aggregate views a lead would actually act on, rather than the biggest set a dashboard could plausibly show.