When a developer hands a task to an AI agent, who actually keeps watching for mistakes?

This explores what happens to the work of checking, steering and catching mistakes when developers hand tasks to AI agents, and when those agents pass work further down the chain. The short version from the corpus: delegation hands off the doing, but oversight doesn't follow it. Oversight moves earlier in time, takes new forms, and some of it ends up owned by no one.

Start with where the effort goes. Interviews with experienced developers show that oversight is not mainly reviewing finished code. It comes in four forms: setting constraints before the agent starts, planning together with it, watching it work, and reviewing afterward. Much of the effort happens before the agent writes a line How do developers actually oversee AI coding agents?. A separate study breaks supervision into seven stages: Plan, Monitor, Wait, Review, Teach, Manual Fix and Update Assets. In it, developers coped with the load by writing reusable guidance files and by having other agents do part of the checking How do developers actually supervise AI coding agents?. So one answer to "what transfers" is that some oversight passes to other agents. What stays with the human is designing that checking: deciding what gets checked, by what, and against which written guidance.

The review that is left is thinner than it looks. Developers rely on shortcuts: passing tests stand in for correct code, a sensible plan stands in for working output, and skimming is the main way of reading Do developers actually review agent work exhaustively?. Anthropic's usage data adds a twist. Experienced Claude Code users auto-approve twice as often as newer users, yet they also interrupt twice as often Do experienced AI users trust agents more or just oversee differently?. Expertise doesn't reduce oversight. It moves oversight from approving each action to stepping in mid-run. One caution: a computer-use study found that planning-based oversight cut problematic actions by 76%, but people blocked only about 21% of risky actions once execution had started, whatever approach they used Does oversight strategy prevent risky agent actions or just catch them?. Setting things up in advance prevents problems. The ability to step in mid-run is often more nominal than real.

People's instincts about what to keep are sharper than you might expect. Trust drops on tasks that can't be undone and that people outside will see, like sending an email. It doesn't drop on tasks that are merely high-stakes but fixable What makes people distrust AI agents they delegate to?. That points to a design principle: the decision about when an agent should stop and ask has to be built into the model, the surrounding software, and the interface, not left to the user's attention How should AI agents balance autonomy with user oversight?. At the organisational level, Deloitte describes the same idea as deliberately placed human checkpoints backed by tamper-proof action logs How should organizations supervise autonomous AI agents?. Everyday users feel the gap. In tens of thousands of Reddit posts, people's values were mostly met by what the agent delivered, but went unmet during supervision Where do user values break down in agent supervision?.

The part you probably didn't know to ask about: once agents delegate to other agents, the person who gave the instruction, the person watching, and the person who gets hurt can be three different parties. Oversight designs that assume they're the same person stop working Who actually bears the risk when multi-agent workflows fail?. When a chain of agents crosses company boundaries, nobody is clearly responsible for deciding whose rules apply. The operator, the organisation, a regulator and a standards body may each have rules, and those rules can conflict or be hidden from each other Who enforces invariants when agents cross organizational boundaries?. Whether a handoff between agents even counts as a violation depends on written policy. Where no policy exists, every exchange is allowed by default How do policies determine whether agent transfers are violations?. So the oversight that transfers is the oversight someone wrote down in advance. Everything else goes unwatched.