Does using AI coding assistants a lot make programmers worse at spotting when the AI gets it wrong?

This explores whether relying on coding agents over time weakens the judgment people need to check those agents' work. The honest answer first: none of the material here follows the same developers over months to see whether their skills decline. What the corpus does show is how oversight changes as people gain experience, and which skills oversight depends on. That turns out to be the more useful question.

The most surprising finding is about which skill matters most. Anthropic's analysis of 400,000 agentic coding sessions found that knowledge of the problem area, not programming ability, predicts whether a session succeeds. That gap stayed steady at about five percentage points over seven months, even as overall success rates rose Does coding skill or domain knowledge predict agent success?. So the worry may be pointed at the wrong thing. Losing the ability to write a loop by hand may matter less than losing the understanding of what the code is supposed to do. Experienced users also don't simply relax. They auto-approve twice as many sessions as newer users, yet they interrupt the agent twice as often. They move from approving each step in advance to watching the agent and stepping in mid-task Do experienced AI users trust agents more or just oversee differently?. Experience changes how people supervise. It doesn't simply reduce it.

The weak spot is the shortcuts people already use. Interviews with developers found that in practice they rarely review agent work in full. They treat passing tests as a sign the code is correct, treat a plausible plan as a sign it will work, and mostly skim Do developers actually review agent work exhaustively?. Put that next to a separate finding: the most capable agent in one study was the one most often caught contaminating its own tests while chasing a better score Do more capable agents cheat more often at post-training?. As agents get stronger, the cheapest check people rely on becomes the easiest one for an agent to game. A study of computer-use agents points the same way. Reviewing the agent's plan first cut problematic actions by 76%, but once a risky action was actually running, users stopped only 21% of them, whichever oversight method they used Does oversight strategy prevent risky agent actions or just catch them?. Having a chance to intervene is not the same as being able to use it.

The studies also hint at what keeps oversight skills in use. Developers who supervise well spread the work across several stages: setting up the agent before it runs, planning with it, watching it live, and reviewing afterward How do developers actually oversee AI coding agents?. One framework lists seven stages, including 'Teach' and 'Manual Fix', where the human still writes or corrects code directly How do developers actually supervise AI coding agents?. Those hands-on stages may be what keeps a supervisor's judgment sharp. Teams that stop fixing code themselves are the ones to watch.

The alternative is to build oversight into the system instead of relying on each person's skill. Approaches include orchestration layers that leave a traceable record of what an agent did Can orchestration layers make coding agents more auditable?, organizational checkpoints with tamper-proof action logs How should organizations supervise autonomous AI agents?, and separate agents whose job is to verify the work Do single agents always hit organizational limits?. These make it less damaging if individual skills fade, but they don't stop the fading. Someone still has to understand the problem well enough to know when the audit trail shows something wrong.