Cognitive Ergonomics (candidate lens)

Research / candidate status

This is a research candidate, not part of the shipped model. The Agentic Maturity model remains five axes and this lens does not affect placement, Spine level, or team rollups. The operational signals below exist in the workspace toolkit today; the scoring bands that would turn them into a maturity axis are uncalibrated and are open work.

What does ship today: aios analyze --push sends a single derived shadow band (ce_band, integer 0–4) team-tier to the Team Brain, where it appears beside AM in the Individual Maturity dashboard — always badged shadow · uncalibrated. The four raw attention signals stay admin-tier, local-only, and never leave your machine. Read this as a proposal with real instrumentation behind it, not a sixth axis you can be scored on.

The five AM axes all protect the work: Verification checks the output, Context hygiene checks what the model sees, Autonomy/leash sets how much you delegate, Learning/compounding checks whether the system improves, Cost & governance checks whether it stays in bounds. None of them protect the operator.

Cognitive ergonomics is the candidate sixth lens: how well a person sustains the attention and judgment that supervising agents demands. As agents get cheaper and more parallel, the human’s attention — not the model — becomes the bottleneck, and the failure modes that human-factors research documented decades ago reappear unchanged.

---

Why it’s a candidate

Three findings from the research wiki explain why supervising agents is cognitively harder than writing code, and why none of the existing five axes captures it:

• The supervision paradox (Bainbridge, 1983) — automation removes the easy work and leaves the human the hard, rare, high-stakes judgments, while eroding the vigilance needed to make them. A developer can be L5 on the Spine and still fail here.
• Attention residue (Leroy, 2009) — hopping between concurrent agent sessions leaves part of your attention stuck on the last one, degrading judgment on the next. Orchestration is task-switching by design.
• Flow and its fragility (Csikszentmihalyi, 1990) — high-quality work needs contiguous focus, which interruptions shatter instantly. Interrupt-driven orchestration is hostile to it unless deliberately managed.

Relationship to the autonomy axis

Cognitive ergonomics is close to, but distinct from, Autonomy/leash. The autonomy axis asks how much you delegate (Sheridan’s levels of automation). This lens asks what that delegation costs the supervising human’s attention and judgment. They are the two halves Bainbridge and Sheridan describe: the delegation scale, and the human cost of climbing it. Keeping them separate is itself an open design question.

---

What it would score

The distinctive claim of this lens is that its signals already exist as operational data — the AIOS workspace toolkit computes them from your own session activity, locally, over active_hours (wall-clock time split into work blocks at idle gaps):

Signal	What it measures
focus_block_avg_min	Mean length of an uninterrupted work block — a flow-condition proxy
context_switch_rate	Project changes between consecutive prompts, per active hour — residue from switching what
interrupts_per_hour	Prompts that hop to a different session, per active hour — residue from switching which agent
concurrent_sessions_peak	Max distinct sessions active in a 5-minute window — supervision load

Today these produce a one-line reading of your attention pattern (deep-work-leaning · orchestration-heavy · mixed · no activity), plus a shadow band (0–4, vs your own baseline) in local reports and — when you push — in the Team Brain beside AM. The band is a coarse mirror, not a calibrated maturity score: we can measure the pattern; we cannot yet say which pattern is “more mature.”

The cognitive-load lens The three failure modes and the full measurement surface.

Protect deep work while orchestrating The shipped playbook that acts on these signals.

---

What’s open (the calibration problem)

Why this is a candidate and not an axis: a maturity axis needs bands — immature → mature — and we do not yet have the evidence to draw them honestly. The hard question is that the “mature” pattern is not obvious. A high concurrent_sessions_peak might indicate skilled L5 orchestration or attention-shredding overreach; a long focus_block_avg_min might be deep engagement or a failure to delegate. The signal is real; its interpretation is not settled.

Calibrating the bands would require:

• Baseline windows — enough per-operator history to separate a normal day from a degraded one, rather than judging against an absolute threshold.
• A decisions corpus — correlating attention patterns with downstream outcomes (were the judgments made under high interrupts_per_hour worse?), so bands map to quality, not just to a preferred work style.
• Guarding against a bad incentive. A naive band (“more focus = more mature”) could penalize legitimate high-orchestration work. The lens must not become a productivity-surveillance metric; it is a mirror for the operator, tier-admin by default and computed locally.

Until that work is done, the five-axis model stands unchanged and this lens stays explicitly a research proposal.

---

See also

• The Maturity Model — the five axes this lens is a candidate to join.
• Research Wiki — the human-factors canon this lens is built on.