Should an AI tool behave differently if you want the job done fast versus if you're trying to get better at doing it yourself?

This explores whether AI tools should offer different controls to someone who wants the job done quickly and someone who wants to learn from doing it. The corpus has no study that compares an 'efficiency mode' with an 'upskilling mode' directly. It does have several pieces that together suggest a design answer: the most useful control isn't a single autonomy slider. It's the ability to decide, step by step, who does each part of the work.

Cocoa is the clearest example Can users steer AI agents better by revising plans mid-execution?. The system lays out a plan and lets users assign each step either to themselves or to the agent, and revise the plan while the work is running. This handles both goals in one interface. Someone after speed hands most steps to the agent. Someone trying to learn keeps the steps they want practice on. Lab users found it easier to steer than a chat baseline and no harder to use, and field users particularly valued being able to delegate work explicitly. The unexpected part is that the controls a learner needs and the controls a busy expert needs turn out to be the same controls, just set differently.

Personal preference may matter less than you'd expect, though. A study of junior and senior engineers found that company policies, such as required tools, allow-lists and data rules, set how much control people keep over agentic AI before individual preference comes into play Does personal preference shape how engineers use AI tools?. Within those limits, novices swung between leaning on the AI too much and avoiding it altogether. That is exactly the group that upskilling controls are meant to help, and they often aren't the ones setting the controls. Designing for learning can't stop at the interface. It also has to reach the organizational defaults that sit above it.

The tool also has to work out which mode the user is in, and agents are bad at this. On UserBench, models fully matched what users wanted only about 20% of the time, and uncovered fewer than 30% of user preferences by asking questions Why do AI agents miss most of what users actually want?. Conversation analysis offers a structured approach: 'insert-expansions', the short clarifying exchanges people use to check intent or agree on scope before they answer When should AI agents ask users instead of just searching?. A question as simple as 'do you want me to do this, or walk you through it?' is one of these. Asking it early stops the agent from quietly doing everything for someone who wanted to learn.

Lastly, the user's goal may change partway through a session. AI context keeps shifting as prompts, history and retrieved material pile up, unlike the fixed screens of ordinary software How does AI context differ from conventional software context?. So a setting chosen at the start isn't enough. The ability to change who does what mid-task, as Cocoa allows, may matter more than any upfront choice. To go further, the open question the corpus leaves is how to tell whether keeping steps for yourself actually builds skill, or only feels like it does.