The hidden tax of agentic coding is context switching

The pitch for AI coding assistants is less typing. The bill is more triaging. Every proposed diff is a small forcing function: read it, evaluate it, decide. The cost is paid in attention switches, not keystrokes. Cumulative switching fatigue, not agent quality, is the limiting factor, and better agents make it worse until the human-in-the-loop model itself changes.

Parent discussions

• Agent-Editor Protocol — “the human stays in the editor, shaping code before accepting” is a stance of constant switching; this note names the cost the spec implies

Connected concepts

• Delegation is not collaboration — names the incentive problem in harness design; this one names its cognitive price
• The smart-editor mode — diffs as a transient interaction mode — the proposed repair: cut the cost per switch, not the number of switches
• Irenaeus — full aliveness as attention — the same phenomenon from the other side of the vault: triage trains you to attend to everything as an inbox
• Stein’s act-content-object triad and the attention economy — formalizes why the “as inbox” mode is invisible to time-on-screen metrics

The point

The mechanism is old and documented. Mark, Gudith, and Klöckner’s 2008 study “The Cost of Interrupted Work” found that interrupted workers finish their tasks (by working faster) and pay for it in higher stress, frustration, and effort. Interruption doesn’t show up as lost output; it shows up as accumulated load. That is the profile of a day with an agentic harness: commits land, throughput graphs look great, and by mid-afternoon you have nothing left. The agentic loop imports interruption as a feature. Each diff is a preemption of whatever mental state you were holding, and the asymmetry bites: the agent’s context is persistent and serialized, yours evaporates on every switch and must be rebuilt each time you return.

Aider’s auto-commit looks like the fix and is the worst version. It defers and compounds the cost. You pay everything at once, later, reading a git history of decisions you weren’t present for, stripped of the context that made each legible. Interrupt-per-diff lets you pay in installments while the state is warm. Auto-commit is a balloon payment.

This is why “agents are getting smarter” answers the wrong question. A smarter agent proposes more diffs per hour. If each diff is an interrupt, agent improvement raises interrupt frequency against a fixed human refractory period. The human is the bottleneck resource, and every gain in agent capability is spent buying more of the thing the human can least afford. The Agent-Editor Protocol’s premise, keep the human in the editor, in flow, shaping rather than reviewing, is aimed at this. But staying in the editor doesn’t abolish the switch; it shortens the resumption. The real fix has to change what a single decision costs, which is the argument of the smart-editor note.

Open questions

• Can switching cost be measured directly in-editor: time-to-first-keystroke after a diff arrives, error rate in the first minute after resumption?
• Is there a batching discipline that beats both interrupt-per-diff and Aider’s compounding balloon payment, or does batching always just trade frequency for depth of context loss?
• Does decision memory (the parse-phase decision log in the protocol) actually lower resumption cost, or merely document what was lost?