Every improvement in this factory made the agents more capable, and every one of those improvements routed more decisions through the same single human, who does not get more capable at the same rate. That is not a complaint about any particular team or tool. It is a scaling law, and a new lab was born specifically to name it and design against it before it turned into permanent overload.
The actual problem
Compute scales. Model quality scales. The thing that does not scale, in a one-human agent fleet, is the human’s attention, and nothing in a typical agentic setup treats that as a real constraint. Agents check in whenever they hit a decision point, the check-ins queue in whatever order they arrive, and the human ends up triaging a stream that mixes genuine judgment calls with things that were never actually ambiguous, just unconfident. The scarce resource is being spent on the wrong kind of question.
What I tried
The analogy that stuck: an air-traffic control tower going up next to an airport that is already busy, where pilots have been radioing one overwhelmed controller at random. The fix is not a second controller, it is standard flight procedures, written once, for the routine cases, so a pilot only radios in when something genuinely needs a human. Mission command is the same doctrine, older than software: give clear intent, then let the people (or agents) executing it act independently inside that intent, and reserve escalation for the cases the intent genuinely didn’t anticipate.
Four supervised agent teams got intent contracts written this way: an objective, what they may never do, what they may do without asking, exactly when they must ask, and what to do if they ask and get no answer in time. All four were ratified as drafted, with no line-edits, which was itself treated as a data point rather than a compliment, because how many patches a first-draft contract needs in its first real week is the actual signal for whether this delegation works.
What worked (so far)
Nothing was built. That was deliberate: the whole design is labeled a hypothesis, cognitive-psychology claims underpinning it were flagged for verification rather than trusted from memory, and a kill criterion was set before day one, two honest weeks of the daily routine, revise the design if it doesn’t help, before any code exists to make the mistake expensive to undo. Day one itself needed a restart, logged honestly as a false start rather than smoothed over, with the bar for a working day lowered to one real window and one debrief rather than a full routine nobody could sustain.
The reusable pattern
Treat the human’s attention as a scheduled, budgeted input to the system, the same way compute or a rate limit would be treated, instead of an assumed-infinite constant everything else gets designed around. Intent contracts are the mechanism: standing intent for the routine cases, explicit conditions (not vague judgment) for what must escalate, and a stated fallback for when the human doesn’t answer in time. The contract’s job is to make silence an instruction, not a stall.
What next
This is a two-week-old lab reporting a design and one real day of practice, not a result. The problem is lived and real; the design is coherent and explicitly unproven, with a go/no-go review scheduled once enough days of honest data exist. The honest version of this post is: here is the doctrine, here is the first day, and here is the criterion that gets to kill it if it doesn’t help.
What Next
Backlog Management for One Human and Several Artificial Interns
The same pillar's earlier answer to the same scaling problem, from the backlog side rather than the attention side: the queue itself as the thing worth designing.