How do developers actually oversee AI coding agents?

Dhanorkar, Passi, and Vorvoreanu interview 17 experienced developers (16 of 17 using agents several times a week for professional work, 12 daily) about how they actually oversee coding agents such as Claude Code and Cline. They report finding "at least four forms of emergent oversight work: a priori control, co-planning, real-time monitoring, and post hoc review," and argue that "oversight work is not only reactive and retrospective, as portrayed in existing research, but also preventative and proactive." A priori control happens "even before prompting," where developers "anticipate failures, preemptively configuring agents to prevent them from making specific mistakes." Co-planning happens before execution, where developers "revise plans with agents and seed solutions to them, pushing agents in desirable directions away from pitfalls."

The paper's reasoning is methodological as much as empirical: the authors used "temporal anchoring to ground participant responses in end-to-end agent workflows," explicitly asking what developers do before, during, and after agent execution, and found that participants organized their own accounts temporally even without prompting — particularly when distinguishing co-planning from real-time monitoring. The authors are careful to note that a priori control looks "less like oversight and more like 'good' software engineering practices," but argue it still counts because participants described doing this anticipatory work "at times with explicit oversight intentions (e.g., preventing known mistakes or failures from occurring altogether)."

This gives a concrete, practitioner-side register to claims that have mostly run the other way in the library. Does granting agents more autonomy undermine human oversight? argues from agent design that oversight capacity degrades with use; this study shows experienced developers responding by pushing oversight earlier, into configuration and planning, rather than relying only on after-the-fact review — a front-loading strategy that the design-degradation paper's excerpt does not address. It also gives empirical texture to What UX principles do workplace users want in AI agents?: where that paper reports stated user priorities around control, this one shows what exercising control actually looks like in practice — deny lists, seeded plans, preemptive configuration — among a specific population of agent power users.

The excerpt does not establish how common or effective these four forms of oversight work are beyond this sample: participants were power users at well-resourced technology firms (12 of 17 at the same large organization), delegating "relatively small tasks," which the authors say likely made real-time monitoring optional and post hoc review easier than it would be for longer-running, higher-stakes tasks. The finding should be read as what proactive oversight work can look like among experienced developers on bounded tasks, not as a general claim about how agent oversight works across skill levels, task sizes, or organizations — the authors explicitly flag longer-running, more complex tasks as open territory for future research.