Capability

AI subscriptions as governed workers

Running vendors' authenticated command-line tools as governed workers, alongside API-keyed and local models

Vendor command-line sessionsAPI-keyed and local open-weight seatsOne governance platform
Three ways to reach a model · one way to govern them One work graph behind every seat
  • 31,811metered usage entries, each tied to the work that spent it
  • 94.5%overall cache hit rate across one priced engineering program
  • 30 sdeadline for a vendor session to confirm a prompt arrived, by its own transcript event

My governance platform runs real interactive sessions of vendors' authenticated command-line tools, under my own logins, as durable governed seats, alongside API-keyed seats and local open-source models served on my own GPUs.

A subscription seat

The vendor's own tool, in a real interactive session under my login. Hooks on that tool turn policy into refusals.

What makes it a governed seat
hosted durably, so a conversation resumes across restarts and account switches
a prompt counts as delivered only when the provider's own transcript shows it arrived
signed events and per-seat cost attribution
the same acceptance rules as every API-keyed and local seat
Reasoning time stays unbounded; confirming that the prompt arrived does not.

Every seat is metered against the work that used it, and cheap is the default: a role contract names its model and reasoning effort as data, so the lead, planning and architecture seats use a heavier model than the development seats on purpose. Start with Directing AI coding agents at program scale and What the work cost, to the penny.

One program, cost by phase · coloured by seat
$227 quality assurance · reviewer · 92% cache
$210 development · builder · 97% cache
$143 orchestration · lead · 97% cache
$55 review · reviewer · 89% cache
$46 architecture · lead · 96% cache
$36 gatekeeping · reviewer · 91% cache
$22 planning · lead · 97% cache
lead seats
builder seats
reviewer seats, another model family
Jobs rendered from the work graph keep stable prefixes, which is where the cache comes from.

Builders and reviewers come from different model families so their blind spots do not overlap.