00 / the concurrency curve
Attention has a peak. The evidence points near four.
Add a second agent and you get more done. Add a fifth and you start losing the thread. Effective output rises with parallelism, then the cost of holding it all in your head takes over. The turn keeps landing near four - and not just for agents. Screens, live conversations, objects you track at once: early evidence points at the same shape, the same neighbourhood. We did not pick four for agents; a century of attention research keeps landing nearby, which is exactly the claim we are now stress-testing.
01 / the curve - effective output vs parallel streams
Pick a domain. The shape holds.
- Domain
- AI agents under supervision
- Peak at
- N = 4partial
- Bands v1
- supervisable ≤4 · saturating 5-7 · overloaded ≥8
- Signal
- Throughput climbs, then the babysitting tax bites. The fan-out plateau sits around eight; your personal peak is what Zeno measures.
Curve shapes here are illustrative defaults. The exact per-domain numbers are being sourced from the deep-research questions below - and, for agents, from your own Zeno sessions. Band thresholds are v1 cut-points read off these curves, published so you can quote and challenge them.
02 / replication - four domains
The same shape, four ways.
AI agents ~4
Supervising parallel coding agents: throughput climbs, then the babysitting tax - context-switching, verification, re-prompting - dominates.
Zeno RTLX-S supervision-load curve · your data calibrates itWorking memory ~4
The number of independent chunks you can hold and manipulate at once. Tracking moving objects tops out in the same place.
Cowan 2001, "the magical number 4" · Pylyshyn & Storm 1988 (MOT ~4-5)Screens ~4
Monitored displays in a control room or trading desk: detection and vigilance degrade as the wall of screens grows past a handful.
Supervisory-control span · vigilance literature (Tickner & Poulton 1973)Conversations ~2
Simultaneous live chats a support agent can hold before response time and quality fall. The floor of active talk is even lower.
Contact-centre concurrency · media-multitasking cost (Cummings & Guerlain 2007)03 / the claim
The claim, stated carefully.
Human parallel-supervisory capacity is not infinite and not linear. Across domains that look unrelated - memory chunks, tracked objects, monitored screens, simultaneous conversations, supervised agents - effective output follows an inverted-U that peaks at a small, strikingly similar number. Our working hypothesis: the peak is a property of human attention, not of the thing being supervised - and R5 below is us trying to break that claim, not just assert it. If it holds, agents are simply the newest stream to hit the ceiling, and Zeno's job is to measure where your peak sits, on your real workload, instead of guessing.
04 / open questions - R1-R6
What we are researching next.
The curve above ships with defensible defaults. These are the open questions that will replace
them with cited, domain-specific numbers - and decide exactly how hard we can make the
cross-domain claim. Full prompts in
docs/RESEARCH_PROMPTS_CONCURRENCY_2026-06-14.md.
- R1 - Capacity ceiling. Cowan's 4 vs Miller's 7±2 vs MOT vs subitizing: what is the consensus peak, and is the curve plateau-then-decline or monotonic?
- R2 - Screens. Where does detection/vigilance performance peak as monitored displays grow (control rooms, SOC, drone operators)?
- R3 - Conversations. The measured concurrency at which live-chat support quality peaks then falls.
- R4 - Span of control. Sheridan supervisory control + Olsen/Goodrich fan-out (fan-out = neglect time / interaction time) mapped onto agent supervision.
- R5 - The invariant. The strongest version of "the peak is a constant of attention" - and the strongest counterargument.
- R6 - Your peak. The curve-fit (inverted-U / Yerkes-Dodson) that locates each user's personal peak-N from RTLX-S + session telemetry.
05 / access - pre-release
Your peak is measurable. Waves are opening.
Zeno measures the supervision cost of your AI-assisted work and renders your personal concurrency curve from real sessions. Access opens in waves during pre-release; waves exist because the instrument is calibrated per cohort, and every early user gets direct access to the researcher.