Copilot often stops after saying what it plans to do next and requires my to tell it to continue

[szymonnowak-st]

Select Topic Area

Question

Copilot Feature Area

Copilot Agent Mode

Body

When using Copilot in VS Code (GPT 5.2 model), it often says what it will do next and just stops. This requires me to type "continue" and then it proceeds with whatever it said it will do. Sometimes, I need to tell it to continue a few times, as even after telling it to continue, it will say again what it's planning to do and stop again. This often happens in the middle of the task, after it has already done quite a few changes.

This makes it harder manage a few agents running locally in parallel, as I need to monitor if they are actually doing anything or waiting for me.

Is it a bug, or should I update its instructions and just tell it to do whatever it plans to do next without waiting for confirmation? It might be a bit risky...

Some examples after it just stops without even asking me for confirmation or anything:

Next I’m going to inspect each uncovered line range in the listed columns and add minimal tests to hit the remaining branches.

Next I’m adding the new keyword management preset hook.

Finding all non-English locale files and adding the new key with the exact same English string.

Next I’ll inspect the menu wrapper and where this menu content is mounted.

etc.

[syedabdulbasitali1]

Hi! This is actually expected behavior for the Agent mode, though it can be frustrating. It is not a bug; it is a "Human-in-the-Loop" safety guardrail.The Agent is trained to follow a Plan $\rightarrow$ Confirm $\rightarrow$ Act loop. When it stops and says "Next I'm going to...", it is effectively pausing to give you a chance to object before it modifies files or consumes a large amount of tokens on a wrong path.However, to reduce the "babysitting" required, you can override this behavior using Custom Instructions.The Fix: Enable "Autonomous Mode" via InstructionsYou can force the agent to be less hesitant by adding a specific instruction to your Copilot settings (or .github/copilot-instructions.md if you have that configured).Add this to your "Code > Copilot > Custom Instructions" (or tell it at the start of the session):"When acting as an agent, if you have a multi-step plan, execute all steps sequentially without pausing for confirmation unless a critical error occurs. Do not announce what you are going to do next; just do it. I authorize all file changes in this session."Why this happens (Technical Context):Stop Sequences: The model often hits a "logical stop sequence" after generating a plan.Laziness/Efficiency: LLMs are optimized to save compute. Generating a plan is cheap; generating code is expensive. The model pauses to ensure the "cheap" plan is correct before doing the "expensive" work.By explicitly giving it permission to proceed without confirmation, you override the default safety prompt that requires it to wait for user approval.

[szymonnowak-st]

Thanks!

I'd expect it to at least ask me if it should continue. Having some kind of system native notification if the current agent session is not in focus would be helpful as well.

I'll try adding these custom instructions and see how it works then.

[syedabdulbasitali1]

tell me if it is working or not.

[gaoyangxiaozhu]

I tried to add rules in my agent md file

and it not work for me , it stop even without hit file update task~~ or when retry tool call~~

[gaoyangxiaozhu]

[gaoyangxiaozhu]

it even stop with multip todo tasks

[tmeindle]

This is not a feature and only happens with GPT 5 models. That's why I stick to Claude.

If I tell it to do a series of steps and it creates a task list it should complete the tasks without me telling it to continue between each task. (which cost a credit each step mind you).

For example telling GPT 5.2 to analyze some attached files and then output a markdown file at a given location with a code review.
Might result in 3 step task - This cost one credit

1. Analyze Content
2. Generate Markdown Content
3. Output the file

while working it will output.
"I have analyzed the files. Starting next step creating code review" but more verbose..

so you have to type "continue" - one credit for the following response
"I have created a draft of the markdown content and need to output it to some/path/that/was/specified.md" and it stops

so you have to type "continue" - one more credit to get the thing you asked for.
"I have created the file at some/path/that/was/specified.md"

So instead of one credit this whole exchange costs 3. The stuff it is outputting is it's reasoning and then it stops and waits for continue.
This is not human-in-the-loop.. Human in the loop should only kick in when it is unsure or trying to do something sensitive like edit a file outside the workspace.

Every Claude model and even gpt-5-mini will do all steps without stopping and produce the file without stopping once.

This is new behavior and it is wrong; The agents shouldn't be designed to nickel and dime us.
It even does it if you tell it to complete every step without stopping

[gaoyangxiaozhu]

agree， in code generation scenario, if we switch use claude or gpt-mini using vscode chat copilot, no this issue happen.
So why it is by design only for GPT-5.2 ?
What's infuriating is that its prompts are also misleading; The GPT 5.2 plans todo tasks but doesn't complete them and doesn't show us whether to continue.

Since you plan tasks you should complete the tasks in the trun or at least If human intervention is required, more explicit instructions should be given, no the one like '**next i will do ~~**'

[gaoyangxiaozhu]

it even stop without multi todo tasks

[gmenounos]

I'm having the same issue with MAI-Code-1-Flash. It will run for a while, implementing some part of the plan but will eventually "finish", displaying "What changed" and "Validation" lists even though it's nowhere near being done. I keep having to ask it again and again to complete all items on the plan.