The agent is confidently doing the wrong thing. Now what?

Ten minutes in, the agent is going somewhere you did not send it. Wrong files, wrong approach, or a careful and completely correct answer to a question nobody asked.

The expensive reflex is to kill it and start again. Almost every one of these is recoverable, usually in one sentence, and the difference between the two outcomes is knowing which of four buttons to reach for.

First, check it is actually wrong

Agents narrate as they work, so the terminal usually says not just what it is doing but why. Read that before acting, because a surprising route is not the same thing as a wrong destination. An agent opening files you never mentioned is very often just working out where things are, and interrupting one that was fine costs you the minutes it needs to find its place again.

What you are looking for is a wrong goal, not a wrong step. It is solving a problem you do not have, editing something you asked it to leave alone, or explaining at length instead of doing. Any of those, act. Otherwise leave it and come back in a minute.

The ladder

The buttons above the message box are in order for a reason: the further right you reach, the more you throw away. Take the lowest rung that works.

• Stop is the mild brake - the agent's own soft cancel, the same key you would press sitting in the bare command-line tool. The turn ends and nothing else changes: the conversation is intact, the agent still remembers the task and everything it read, and your files are as they were a second ago.
• Interrupt is the hard version, the one a terminal has always understood. Reach for it when Stop bounced off, not before. Neither asks you to confirm; both are meant to be pressed the instant you decide.
• Compact is not a brake at all - it is the answer to a full context window. It asks first, summarises the conversation so far and frees the room back up, keeping what the session has learned.
• Clear context resets the conversation in place. The process keeps running, same session, same card; what goes is everything you have said to each other.

Most steering ends at the first rung

And it is worth seeing why. After Stop, the conversation is whole - so do not re-explain the task from the beginning. Correct it, and name both halves: what to abandon and what to do instead.

"Stop rewriting the parser - the job was only to add the new date format, leave everything else alone" lands. "No, not like that" does not.

Then read the first minute of what comes back, because the quality of your correction decides whether you are standing here again in ten minutes. If two corrections in a row have not landed, stop correcting. At that point the conversation itself has become the problem, and that is what the other two buttons are for.

When it is not going wrong, it is just full

On agents that can report it, the right-hand end of the bar carries a small gauge showing how full the agent's working memory is, re-read every few seconds. It runs green, turns amber at seventy percent, and red above ninety.

This matters because an agent whose context is nearly full starts forgetting the beginning of its own task, and that looks exactly like an agent going wrong. If the gauge is red and the work is otherwise fine, compacting is the whole intervention. It can take a minute or two, and nothing is sent to the session afterwards - it will sit there until you say something.

Clear context is for the other case: the agent misread the task early and everything it now believes is built on that. Summarising would only carry the misreading forward in fewer words. That is a residue, not a moment, and the only fix is to start the conversation again with the session still standing where it is.

Each button, what it costs you and how to tell which one you need: steer a session that is going wrong. The same controls from the Director's side: starting and steering agents.