Cognitive Load

When you supervise several agents at once, the bottleneck stops being the machine and becomes your attention. This lens collects the three human-factors failure modes that show up most reliably in agentic work, and connects each to a signal the workspace toolkit already measures.

---

Three failure modes

Attention residue

Switching between tasks leaves part of your attention stuck on the one you left — measurably degrading your judgment on the one you moved to (Leroy, 2009). Round-robining between concurrent agent sessions is high-frequency task-switching by design, so residue accumulates: your review of session B is worse because session A is still resolving in the back of your mind. The cost is on quality of judgment, not just time.

Vigilance decrement

Humans cannot sustain attention on a source that rarely needs them; effective monitoring degrades within tens of minutes (Bainbridge, 1983). “Watch every diff and step in when it’s wrong” is exactly this failure mode — the agent is competent 95% of the time, which lulls you into low vigilance right before the 5% that needs you. Passive monitoring is the wrong verification strategy for the same reason it fails pilots and plant operators.

Supervision fatigue

The compound cost of the two above, plus the shift Bainbridge described: automation removes the easy work and leaves the human the hard, intermittent, high-stakes judgments. Orchestrating agents is tiring in a way writing code is not because the residual job is dense with exactly those judgments, delivered on the agents’ schedule rather than yours.

---

The measurement surface

These failure modes are not just felt — the AIOS workspace toolkit instruments them. The analysis layer emits a set of attention (sanity) metrics from your own session activity, computed over active_hours (wall-clock time split into blocks at idle gaps):

Metric	What it measures	Failure mode it proxies
focus_block_avg_min	Mean length of an uninterrupted work block, in minutes	Flow conditions — short blocks mean an interrupt-shredded day
context_switch_rate	Project changes between consecutive user prompts, per active hour	Attention residue from switching what you’re working on
interrupts_per_hour	User prompts that hop to a different session, per active hour	Attention residue from switching which agent you’re supervising
concurrent_sessions_peak	Max distinct sessions active in any 5-minute window	Supervision load — how many threads you held at once

What these are — and aren't

These metrics ship in the AIOS workspace toolkit’s analysis layer today. They are a described reading of your attention pattern — deep-work-leaning, orchestration-heavy, mixed, or no activity — computed locally from your own session logs. They are deliberately not an AM axis score: the cognitive-ergonomics lens that would score them is a research candidate, and its bands are uncalibrated.

Privacy boundary: the four raw signals stay on your machine (admin-tier). aios analyze --push sends only a single derived shadow band (ce_band, 0–4) team-tier to the Team Brain — badged shadow · uncalibrated in the dashboard, never entering radar/spine/rollups. The metrics exist; full calibration is open work.

---

Why measure attention at all

Because the other AM axes protect code quality, and none of them protect the operator. Verification checks the work; context hygiene checks what the model sees; the autonomy axis checks how much you delegate. But a developer can be excellent on all of those and still be running so interrupt-shredded a day that their judgment — the thing every one of those axes ultimately depends on — is degraded.

Measuring attention makes the invisible cost visible, and gives you something to act on: fewer concurrent sessions, longer focus blocks, batched interruptions. The playbooks turn those levers into workflows.