Osmani argues that adding parallel AI coding agents does not scale a person’s output linearly, because the agent does the generating but the person still does all the evaluating, deciding, trusting, and integrating — work that stays single-threaded no matter how many agents run at once. He calls the point past which supervision quality degrades a person’s parallel agent limit, or ceiling, and argues it is not a fixed number: it shifts with the complexity and novelty of each thread’s task, with how clearly the work was specced upfront, with how long a session runs, and with a person’s own state on a given day.
Usage
He reports his own ceiling as roughly three to four threads for a typical session, reached by treating long agentic sessions the way he treats deep-focus work: defining a session’s duration and scoping each thread to something resolvable and reviewable within it before spawning any agent, since time-boxing creates checkpoints that bound vigilance rather than leaving it open-ended.
When It Applies
Osmani frames finding this limit as a skill mostly learned the hard way — by running past it and mistaking the resulting state, in which output is accepted without being reviewed carefully, for productivity. He offers a calibration heuristic for finding it deliberately instead: start one thread below what feels comfortable rather than raising the count until something breaks; watch review-quality confidence rather than agent count, since dropping confidence in what is being accepted is a more honest signal of having reached the ceiling than any rule of thumb; treat rising anxiety appearing across more than one thread at once as an early warning; and prefer reducing the scope of each thread over reducing how many threads run. This is presented as the author’s own reported practice rather than an established or measured convention.
Related Terms
Ambient Anxiety Tax, Cognitive Debt, Orchestration Tax, The Code Agent Orchestra - what makes multi-agent coding work