How do developers actually supervise AI coding agents?

A qualitative study of 19 experienced developers reconfigures Sheridan's framework of human supervisory control — originally five stages (Plan, Teach, Monitor, Intervene, Learn) developed for supervising automated systems such as aircraft autopilots and industrial robots — into seven stages specific to AI coding agents: "Plan, Monitor, Wait, Review, Teach, Manual Fix, and Update Assets." The paper observed participants working on their own coding tasks with their usual agent setups, then asked them to redraw their supervision workflow as a diagram. Applying the resulting framework to public developer discussions on Reddit, the authors found that "supervisory demands extend across stages," meaning the work of overseeing an agent is not concentrated at one point in a task but distributed across planning, watching, waiting, reviewing, correcting, and asset-building.

The paper's account of how developers absorb this spread is given in its own terms: developers manage the demand "by concentrating effort in planning, delegating supervisory work to other agents, and turning recurring guidance into reusable assets" such as instruction files, tests, and hooks. Its design-implications section extends this into two directions for builders: agents should be designed to preserve developer intent and constraints over long executions rather than drift from them (reducing the Monitor/Intervene burden), and to reflect project- and developer-specific quality preferences beyond functional correctness, since developers still had to fix code that was "repetitive, overly defensive, or difficult to maintain" even when it worked.

This gives empirical structure to what Does personal preference shape how engineers use AI tools? found about who sets delegation depth: where that paper shows policy decides how much agency developers keep, this one shows what happens procedurally once agency is delegated — the seven stages are where the kept-back supervisory work actually lands. It also gives a named home to one of the coping strategies the Dr. Claw system in Can orchestration layers make coding agents more auditable? operationalizes as a built system: "delegating supervisory work to other agents" is exactly the Monitor/Review-stage move that orchestration layer automates. And it complicates Does granting agents more autonomy undermine human oversight? by breaking "oversight" into seven distinct moments rather than one undifferentiated capacity that erodes uniformly — a design fix aimed at the Plan stage (clarifying goals and scope up front) does different work than one aimed at Manual Fix (correcting code after the fact).

The study window was short — about 40 minutes of observed coding per participant, extended only by a retrospective whiteboard exercise asking developers to generalize to other tasks — so the framework describes a structure of supervisory moments, not a measured distribution of how much time or effort each stage actually consumes over a multi-day or multi-week project. The Reddit analysis that follows is described only as confirming demands "extend across stages," without a count of posts or a quantified breakdown by stage. And the design recommendations (preserve intent, align with quality preferences, support multi-agent delegation) are implications the authors derive from the interview and diagram data, not interventions the paper tested; whether they actually reduce supervisory load for developers remains open.