The situation
Every system in this room runs on top of AI sessions doing the actual work, and at some point the way those sessions were organized became a system of its own, worth designing on purpose instead of letting it accrete. The original approach was one long-running session per project. That fails quietly: a project pin goes dead the moment the project pauses, while the type of work behind it (research, building, anything carrying a partner's name, anything involving numbers) never goes stale, because there is always more of it.
So the firm redesigned around stations by type of work rather than by project. A station holds its configuration (which tool tier it runs, how much it is allowed to think, what it can touch) permanently; the conversation inside it does not need to survive and is not meant to, because the actual work product lives on disk, not in the transcript. Alongside that, the firm measured what a session actually costs before a single word of new work gets typed, and built a review chain that any unattended, scheduled change has to clear before it ships.
One long-running AI session per project goes stale the moment that project pauses, while the underlying type of work never does; and nobody had ever actually measured what a session costs just to exist, before any real work happens inside it.
Redesigned around a small number of durable stations, each configured for a type of work rather than a project, with the rule that a station's job is to end by naming what it produced, where that lives on disk, and which station is next; measured the fixed cost every session pays before any work starts; and built a mandatory review chain, including an independent adversarial pass, before any unattended scheduled change goes live.
A defined, small set of stations replaced a growing pile of project-specific sessions. The fixed cost per session was measured directly rather than assumed. A firm rule now blocks any unattended change from shipping until it clears a multi-step review chain, including a check that did not write the change reviewing it independently.
Read on
The decision and the reasoning
Why we did it this way, told first.
ReadWhat we did and what it produced
The work, decision by decision, and the result.
ReadA slice of the project list
A few related projects.
ReadPrivate side
The full report, the source files on record, the timeline and every figure with its source. Password.
Enter