[Feature Request] Ability for coding agent to ask clarifying questions before starting to work

[BOLT04]

Describe your question

It would be great to have some setting or toggle for Copilot coding agent to ask clarifying questions before starting to work. This could be used only sometimes, e.g. on complex issues.
I think this is often in many prompts when using Copilot in chat mode or other features for example.

Example

Prompt for Copilot coding agent:

If you need more information before starting to work, ask as many clarifying questions as you need to complete the task successfully.

Copilot comment:

I'd be happy to help implement this issue. Let me ask some clarifying questions to better understand the codebase and the problem before proceeding. Here are my clarifying questions before implementing:
1. The issue mentions XYZ...
1. Should this apply to X?
1. ....

Once I have these details, I can start implementing.

User comment:

1. Yes, XYZ is the file you should look at
2. No

Then the 👀 reaction is used on that user comment to signal that the agent will start working, using the comments as context, just like it does today.

Motivation
This feature could help us give more important details to Copilot that we didn't have in the issue description. In theory this could improve Copilot's output in the first draft PR, instead of the user only giving feedback to Copilot once that draft PR is done.

Does anyone find this useful to have as a new feature for Copilot coding agent?
Thanks!

[nhu-do]

Thanks for sharing this feedback and idea @BOLT04. This is certainly something that we've also been thinking about. Question for you - the agent represents an asynchronous and autonomous experience with some human intervention on critical portions like merging. In your workflow, would you prefer:

• Using natural language to initiate this "mode"
• Having options for the agent to pause for X seconds prior to execution
• Something else?
  Curious to understand how you imagine this working in the ideal world 😄

[BOLT04]

thanks for your response @nhu-do 🙂 , yeah the "Using natural language to initiate this "mode"" seems better. Now that I think about it we could leave a comment in the issue with that prompt, no need for a specific setting maybe:

from what i understood comments are pulled in context for the agent, just has less "weight". People could save this in their "saved replies" on GitHub, then copilot responds like my example above, etc. I tried it too since I didn't know if we could reference/tag copilot in issues, but it didn't work. I didn't find in the GitHub docs, but tagging probably only works in a PR.

Anywho, even having the trigger in comments with that prompt doesn't seem ideal for:

• users who use coding agent from an IDE (e.g. vs code), or Agents panel, or MCP (i believe people can trigger coding agent with the create_pull_request_with_copilot)
• a scenario where the conversation is back-and-forth and this can basically mean more tokens going to context -> possibly worse performance.
• prompt optimization, where this prompt would be better done by GitHub engineers, rather then inputted by users in a comment. Although this gives more control to users

Hmmmm maybe this "clarifying questions" step would happen "somewhere", then the result of that chat would be added as a comment, or at the end of the issue description, to be used in context for the agent.
wdyt?

[nhu-do]

That's right, tagging Copilot in other areas than PRs won't work at this time, although you're getting to an interesting experience there of having the agent be able to help you curate a plan that the an agent could understand and execute well on.

Agreed completely that this is something that we can better tune the agent to do to build in this sort of chat interaction during the different planning or thinking phases of the agent itself. Something like this can improve the agent's experience without wasting several rounds of iteration and time later down the line.

Thanks for sharing your thoughts, you've caught us at a great time where we're starting to flush out the end to end experience for something like this 🙌🏼