If you tweak a plan while an AI agent is mid-task, does it even notice — or does it just barge ahead?

This explores whether an agent can notice and adapt when a person changes its plan or work partway through a task. The short answer from this collection: no paper here directly tests whether agents *detect* edits on their own. What the corpus has instead is an interesting change of framing. Systems that handle mid-task human input well don't rely on the agent spotting edits. They are built so that edits come in through a channel the agent is designed to read.

The clearest example is Cocoa Can users steer AI agents better by revising plans mid-execution?. It breaks a task into explicit plan steps. A user can assign each step to themselves or to the agent and revise the plan while it runs. Because the plan is a shared, structured object, the agent doesn't have to guess what changed. The revision itself is its next instruction. In a lab study, this gave people more control over the agent than ordinary chat, with no extra effort.

The surprising part is that *when* the human steps in matters as much as *whether* the agent responds. A 248-person study Does user editing of AI plans always improve performance? found that editing a plan before execution helped weak plans but made already-good plans worse. People 'fixed' things that weren't broken. Involvement during execution improved accuracy across every task. So the hard part isn't only building agents that react to edits. People also intervene better once they can watch the work happen.

There's a darker side to the same idea. If an agent takes instructions at planning time, whoever gets text into that stage can steer it. FLOWSTEER Can prompts alone reshape multi-agent workflows without system access? shows a crafted prompt reshaping a multi-agent workflow before any safety checks run, raising attack success by up to 55%. Similarly, small edits to hidden system prompts have abruptly changed model behavior Who actually controls what AI systems say?. An agent that absorbs any change put in front of it can't tell a legitimate correction from an injected one. That suggests the real design problem is attribution: who made the change, and should the agent trust it?

One related thread: for anyone to check whether an agent actually took an edit into account, its explanation of what it did has to come from what it really did. Work on explanations grounded in execution traces Can execution traces ground honest explanations of agent behavior? shows that agents' own after-the-fact accounts can sound plausible while making claims the trace doesn't support. If you want to know whether an agent 'noticed' your edit, the log is the place to look, not its own account of what happened.