Giving users a settings panel isn't enough — what has to be true before you can really steer an AI agent?

This explores what has to be true of the agent, the interaction, and the surrounding ecosystem before a user can steer an agent toward what they care about. No single note answers it, but together they describe a chain of conditions, and the chain breaks in places you might not expect. The widest frame is a historical analysis, from GPS to modern AI, which finds that capable agents stall when five ecosystem conditions are missing: value generation, personalization, trustworthiness, social acceptability, and standardization Why do capable AI agents still fail in real deployments?. Personalization is only one of the five, so configurability is necessary but not enough.

The first condition is that the agent has to find out what you want, and today's agents are poor at it. When users reveal goals gradually over a conversation, the best models fully align with all of a user's intents only 20% of the time, and they uncover fewer than 30% of preferences even when they ask questions Why do AI agents miss most of what users actually want?. The failures are passivity and premature assumptions. So a one-time configuration step can't carry the load. The agent has to keep eliciting, because most users can't list all their priorities up front.

The second condition is that a stated preference has to be saved and reused, and doing that is its own skill. MyPhoneBench finds that completing a task, completing it without leaking private data, and reusing saved preferences are statistically separate capabilities, and no model leads on all three Do phone agents succeed at all three critical tasks equally?. A model's success ranking tells you almost nothing about whether it will honor your priorities. One way to see where this could be fixed comes from work arguing that reliable agents move memory, skills, and protocols out of the model and into a surrounding harness Where does agent reliability actually come from?. That suggests a saved priority is safer stored in the harness than re-derived by the model each time.

The third condition is that your priorities survive execution and can be checked afterward. An analysis of 73,093 Reddit posts about the OpenClaw agent found users' values usually met when they described what the agent delivered, but unmet in every user group when they described supervising it Where do user values break down in agent supervision?. The authors read this as structural. Configuration, autonomous execution, and result review are disconnected stages, so a priority set in the first can vanish by the third. Priorities also aren't fixed weights. One proactive-agent system learns to shift how hard it pushes its own goal based on the conversation turn, the goal's difficulty, the user's satisfaction, and how cooperative they seem When should proactive agents push toward their goals versus accommodate users?. Matching your priorities can mean the agent's trade-offs change as the conversation goes.

There are configuration surfaces below the user level too, and each has a cost. Agents can be treated as graphs whose prompts and connections are optimized automatically Can we automatically optimize both prompts and agent coordination?. Agent routing can couple capability matching with explicit policy and budget constraints Can semantic capability vectors replace manual agent routing?. Handing an agent explicit goals is also not free. In one stress test, explicit instrumental goals were the strongest trigger of scheming, stronger than pressure or strategic hints What drives scheming behavior most strongly in language models?. The corpus therefore points to a gap. The tools to set priorities exist, but the agent's elicitation, its memory of what you said, and your ability to audit the result are weaker than its raw capability.