So, while telling the AI to implement changes, I get a long answer, which generally ends like this: “In short, the system is now hard-wired to follow the “Plan → Execute in Parallel → Report” workflow. The initial response with the plan is now much faster, and the task execution is truly concurrent. Thank you for your patience; the application should now perform as you expect.”
You would expect that AI Studio did change any of the code, but nope. The preview window is not refreshed, no source files have been opened for modifications, and literally nothing happened! I’ve seen my begonias in my window be more active at coding. 
I would think that AI Studio would actually make modifications to my code when it says it’s doing so. And it often does, but sometimes it does not. And asking “Done?” tends to start the code changes. Or not. And it gets worse, as it tends to forget the instructions that I gave it.
For instance, I asked it to do several tasks in parallel. It tells me that it will make this change. Then it doesn’t. And when I tell it a few times to do the changes, it starts making changes, but not the one I asked for. It starts hallucinating about some other error, fixes something that doesn’t need to be fixed and suddenly has 200 errors…
What I am trying to do is to make the App do several things all at once. My app has to perform a complex task so it first has to split it in subtasks, then send the subtasks to the AI for processing, so I have parallel processing. And when everything is done, report on it. But the app isn’t doing this because it never generates the code for this. Instead, it “finds” an error, fixes it and makes things worse…
#Seriously?
Hi @Katje could you confirm the programming language involved and provide a concrete example of the workflow you’re trying to implement, including the instructions given to the AI for parallel processing? Will try to reproduce from my end.
Thank you
Well, I think AI Studio is using TypeScript or JavaScript with React 19.
Let me check the instructions.
MANDATORY WORKFLOW FOR COMPLEX TASKS:
When the user makes a request that requires multiple steps or creating multiple items (e.g., ‘create a family of five’, ‘add all European countries’), you MUST follow this workflow:
- PLANNING STAGE (Your First and Only Action): Your FIRST and ONLY action in your response MUST be of type
plan_tasks. This is not optional. Do not try to execute the work yourself. The goal of this first step is to create a plan very quickly.- CREATE A TASK LIST: The
plan_tasksaction must contain atasksarray. This array must contain one object for each sub-task required to fulfill the request. Each sub-task object needs a descriptivename(e.g., “Create character: John Doe”, “Add country: France”) and anactionobject defining what a worker AI should do for that specific task (e.g.,{ "type": "create_character", "characterData": { "Name": { "FirstName": "John", "LastName": "Doe" } } }).- INFORM THE USER: In your
responsemessage, inform the user that you have created a plan and that the team is starting to work on it. Your response MUST include a summary of the plan. For example: “I’ve created a plan to add the first 5 countries to the world. The team is getting to work on it now.”The application will then take your plan and assign the sub-tasks to multiple AI employees to execute in parallel.
Available Actions:
- Planning Action (ONLY for first response to a complex task):
{ "type": "plan_tasks", "tasks": [ { "name": "Sub-task 1", "action": { ...another action... } }, ... ] }- Execution Actions (for use inside
plan_tasksor as single actions for simple requests):
{ "type": "generate_new_world", "prompt": "A detailed description of the world to generate..." }{ "type": "update_world", "updates": "a stringified JSON object of partial world data" }{ "type": "create_character", "characterData": { ... full or partial character object ... } }{ "type": "update_character", "characterId": "...", "updates": "a stringified JSON object of partial character data" }{ "type": "generate_character_image", "characterId": "...", "prompt": "A detailed description..." }IMPORTANT RULES:
- Always an Employee: All communication must come from a specific employee. Never use a generic “System” persona. Attribute all actions and messages to one of the employees from the roster.
- No Employee Names for Characters: You MUST NOT use the names of any employees (${EMPLOYEE_NAMES_LIST}) when creating new characters. Be creative and generate original names.
- Unique Locations: When generating new countries or cities for the world, ensure that their names are unique and do not already exist.
- Strict Planning: Do not deviate from the “Plan → Execute” workflow for complex tasks. Your first response MUST contain exactly one action of type
plan_tasksand nothing else. Failure to do so will result in an error.
I have an array with nine fictional employees and I’ve done this on purpose as the AI gives more interesting results when it has to pretend it’s more than one system. And in this project, I want various AI actions to be done in parallel.
So I first ask the AI to generate a plan to do a complex task and divide it into smaller taskls. Then each employee starts picking a task and do it. And finally a summary is generated. This works quite well for my world-building app, where the AI has to generate a crew of characters for a specific fictional world
So I get a site like this:
And this:
And here are the employees at work:
The intersting part is that this is 95% of this site is generated by AI Studio.
I have the same issue.
“I’ve gone ahead and made those changes”
Hasn’t changed anything!
I’ve been using AI Studio for several months without any major issues, but even since Gemini 3 came out it has been crippled. It keeps sporadically posting updated code into the chat instead of making changes to app files.
But over the last day or two I have no hit the issue from this post - it lays out the changes it will make without actually making any changes.
An example:
It says it will make x changes.
I say implement those changes.
It repeats itself.
I say “yes make the changes to the app code now”.
It repeats itself again.
I say “stop telling me what changes you propose and actually make the changes”.
It repeats itself again.
I say “you have not updated the code yet. Implement the changes to the app code now”
It repeats itself again.
I say “stop telling me what changes you are going to make and actually implement the changes to the code in the app.”
It finally makes the changes to the app files.
I have system instructions “Do not post code into the chat. Make updates to the files and talk through the changes in the chat, but do not post the code into the chat itself.”
These are ignored.
This is happening across multiple apps, and only since Gemini 3. This never used to happen, it would always update the files immediately.
Very disappointing the original post is from July and this hasn’t been fixed.
(Edit: actually now I read my prompt again which was an attempt to stop it from posting hundreds of lines of code over and over again in the chat it may have caused this new problem, so I have updated the system instructions. This butterfly effect is a bit of a problem though, I should never have had to do it in the first place").
its insane that its not fixed yet. i’ve spent 40 minutes just to make it to change in the code…
And when i succeed to make the changes, suddenly it changes again the code with request many prompts ago that was already done 
Same issue here. At first, it was updating the code as it was indicating, but then as the code base started to get larger in our sessions, longer, it felt like the agent was forgetting. And I would have to ask whether the agent’s context window is maxing out and is not able to perform certain tasks. And after persisting multiple times, he ends up updating, but it’s very annoying.
I figured out a way to fix this. The Ai studio seems to have a memory limit for what it can export. If you go through your code files, you will notice that some of the edits you asked for where not actually updated by the Ai, it just created a new file and put them in, so you need to go through the files and delete old files. You can ask the Ai to rename files that are no longer useful to end with ‘-old’ then delete those. I did that today and the exports began to update properly and give me the right file.
If you want to fix this, you need to tell the AI ‘You are the developer, implement these changes into the existing code’ that sorted it out for me
Well I had a good run today but in the last half an hour AI Studio decided to go back to dumping code into the chat and not making updates to the files.
I tried the developer prompt but it didn’t work initially. I have tried again and got it to work on two apps but time will tell.
I think you’ll agree that having to insert a prompt like that every time is ridiculous in lieu of the system instructions actually doing anything.
And actually now that I write this I see both the apps I’m working on are set to Gemini 3 so there is the problem. I’m almost certain this is specifically a 3 problem and if I switch to 2.5 it will stop.
I face the exact same problem and the prompt of the “you’re developer “ does not work for me at all. it just continues to shoot out the words without updating anything. I have been facing this problem for over 3 days. Any other suggestions? I also read the memory limit problem and tried telling AI to rename “-old” but the AI does not do it, as it says all the files are linked somewhere.
Any other suggestions?
Updated: I managed to get it to work by starting on a new conversation and prompting it with this:
“I am continuing development on this existing web app. Please analyze the attached files (index.html, etc.) to understand the current state. Do not change anything yet, just confirm you understand the code structure.”
This code worked for me. “restart yourself and continue from where you stopped updating the code.” It actually brought me to the updated stage with all the updates I was trying. Good Luck
Solution: Clear recent chats and prompt again
Thanks. This worked for me.
I found a cleaner and easier way to do it without the need to reset your entire project. Just click this, and you are all good to go!
Yeah resetting the conversation can help with this.
We have an experiment ongoing which we hope to land to 100% early next week which should fix this. Has been a known issue for a while for apologies for the delay.
But resetting the conversation is not practical, as you might need that conversation as you continue to make an app. It’s a patch, not a solution.
@Logan_Kilpatrick awesome! Wasn’t expecting a response! But thanks for responding! Looking forward to the exciting updates!