The human attention bottleneck

Roughly three concurrent agent sessions per person. One genuinely complex task at a time — with three complex features in flight, the first is great, the second mediocre, the third rubbish. Deep systems thinking spent after a single deep task. And irreducible roles like on-call that set a floor on team size no agent can fill: a primary plus a secondary, with nobody on call more than half the time, gives a minimum viable team of four.

The first automation that opened a pull request for every high-impact finding failed — nobody wants eighty pull requests they will never read, the familiar graveyard of seven hundred open issues. The pivot is the design rule: convince the human it is worth the attention, instead of convincing the agent it is worth the tokens, because the agent will always think it is. That means scoring, compressing, and surfacing a small number of chosen actions rather than emitting everything the system could produce.

Speeding up one stage of an unchanged pipeline is Amdahl’s law applied to delivery: accelerated engineers queue on whatever stage nobody rebuilt. One organization fixed AI fluency across the whole company and still saw no velocity gain in the following quarter, because agent-accelerated engineers sat idle waiting on waterfall product handoffs. What was a smooth path from idea to production becomes a car crash somewhere after the code is written.

Cognitive load is the sleeper issue — think very hard, ten minutes, think again, with no relaxing coding stretch; one customer sends a person home for a week every few weeks just to decompress; people nap after three hours of agent supervision. The variable-reward loop of watching commits land at 11pm is described, repeatedly and independently, as addictive.

Also: The Goal (theory of constraints); the DORA 2025 report; Multitudes on agentic-adoption overtime.

Assay’s desks run silent by construction: state lives in claims, the roster and pull requests, and anything needing a human is a filed issue rather than console output. The direct implication of this evidence is that a regression which surfaces console noise or demands ambient attention is a severity bug, not a polish item. Assay also sits at the middle path the open-stack argument asks for: the gate is one human decision per pull request at the merge boundary, not a per-command approval prompt.

Authoring and review are accelerated; the queue that visibly grows is human merge and ready-flip capacity. That is a process-redesign problem, not a patience problem, and no redesign is currently designed. Two mechanisms are proposed and unbuilt. First, risk-and-size triage labels attached at verdict time, so a single merge gate can be worked in priority order rather than arrival order — the auto-labeler idea applied to Assay’s own queue. Second, instrumenting four stage rates — briefs authored, briefs dispatched, pull requests reviewed, evidence rows verified, per week — and managing to whichever queue is growing, which turns “back pressure” into a measurable signal about where the next role or human hour belongs. That is how it runs territory, and it is not instrumented yet.

The minimum-viable-team floor — agents cannot hold the pager — gives Assay a crisp boundary: agent fleets own authoring through merge; humans own production consequence.

• Whether anyone publishes a review-capacity number — pull requests per reviewer-hour at a stated quality bar — rather than the current anecdotes (3 sessions, 12,000 lines, 80 PRs).
• Whether decision-queue ergonomics becomes a product surface: batched, evidence-pre-digested, one-keystroke accept/reject is described by nobody in this scan as shipped.
• Whether the cognitive-load reports attract any measurement, or stay testimony; a burnout signal with numbers would change how fan-out is scoped.