One engineer, a room full of agents, and a single discipline running through every project: let the machine do the work, but make it prove what it produced before anyone acts on it. An engineer's working notebook: each method below is explained, animated and counted.
Those four steps run on AWS, in this lab’s own backend, and on the agent factory that builds the work.
counted by the machine itself · nothing on this strip is typed by hand
each with its own checkable receipt · press the buttons
Agents that merge on their own word are a liability.
A spec goes in; a reviewed pull request comes out. Gates, multi-model review, a human on approvals, not in the loop.
A model does not judge its own family.
The reviewer seat always belongs to a rival vendor, measured on production data rather than assumed.
Fifteen agent sessions and a factory, one person.
One coordinating session dispatches to named peers, opens specs and launches rounds, and verifies what comes back against disk and git rather than against the report.
Capability without restraint is a demo, not a system.
Autonomy is earned in tiers: reading is free; spending waits behind a gate; the seat that pages a human never opens at all.
Every quarter the models get better and the prompt gets rented for less. What does not commoditise is the layer that makes a generated answer safe to act on: the substrate a system accumulates, and the verification that lets someone use its output without checking it by hand.
I don't build assistants that sound confident. I build systems whose output you can ship without re-reading it, because a deterministic gate already did.
Years of engineering where a wrong number costs money taught me what one costs. Two years building agent infrastructure taught me how to stop one from reaching a human. Every system here sits on that seam: the place where autonomy meets accountability, and has to be earned.
the summary register · every line checkable · each lives on its method card above
A frontier model in every seat of a multi-agent committee made each agent flawless, and changed the group’s behaviour by nothing (+1.7 pts, p=0.62).
One model family reviewing its own code: 0% blocking verdicts. A rival vendor on the same code: 43%.
A spec goes in; a reviewed pull request comes out. ~140,000 lines of TypeScript, 794 specs, one person.
A melody hummed into a phone at night comes out as a mastered track with a video cut to measured beats. 108 songs, 164 masters, about 10.5 hours.
where the methods run · context, not the showcase