Thanks for sharing a roadmap… just hope it wasn’t in order of priority or ship date because whats last should be first!
This will be great, especially if there is a button to press that would be something long like ‘explain this’ so after an AI puts together something it can then explain each piece and rationale in a sort of interactive onboarding way, with highlights and explanation features.
I think if bubble uses it more so it doesn’t offload cognitive functions of the user, and instead builds a user real understanding of the platform, that will be really powerful.
Thanks for your response. To deliver those bullets, are we talking months or years? Just to get a feeling about how hard it simple those points are.
When are we able to attach documents, screengrabs, or full app docs to the ai to allow it to do a full app. Also allow unlimited text in the prompts. At present it is extremely limited to define a full schema for the app so it can properly create a detailed full stack app. Having tested the last week or so I find it can no way complete an app, even the functionality it produces is not complete, ie filtering repeating groups, galleries and prompts to change layouts. It is miles away from, what you prompt. It maybe only good for layouts for now. It should also access the bugs so it can fix.
Whats ur setup like? LLM? plan b4 act? How do you manage ur context window? Your Markdown files?
I’m able to get nearly perfect JSON… if only I could get newly generated ids without using the private api calls I’d have really a useful AI agent to build Bubble apps. Which all the more makes it frustrating that Bubble’s agent has so little utility.
Would you pay for a course that teaches you how to get it right? Do you think the market for such a course is large?
Asking because I am not sure if I would put time in it or not. It seems there are two camps. One camp is pro ai and eager to figure out how it works and delivers solid code. The other camp keeps repeating how bad lovable and AI coding in general is. This second camp does not seem to be interested in changing their opinion. The first camp just puts in the hours, improving by doing and find out themselves what works and what not.
My project is now close to 100k lines of code and every 25k extra lines or so meant that I had to redesign my guardrails. One of the most valuable decisions you can make is setting up a powerful logging system. In my system with opentelemetry on and single logging db that merges container logs, webserver logs and application logs, which AI can access by using a rich api or simply cli, it will see any error and autofix it. Requirements both business and non business are relatively easy. I let AI write them, I review them and then AI writes test scripts, then writes the code, then test them and handoff to me for final testing. Now working on Husky and more complex linting such that most guardrails are enforced without me needing to worry about it. It is getting closer and closer to a state where I do not have to spend time thinking about technology but rather “speak” to the system using only business language. Btw…Google real time AI speaks so natural and latency is so low, that I do think that the interface will become indeed speech rather than written text in the foreseeable future.
I just use the ChatGPT paid plan interface. Lots of long term memory stored on how I like responses to be. For the most part it works fine to provide correctly structured response format, but sometimes it doesn’t.
What I’ve noticed is that it often enough in coding tasks ignores clear and explicit instructions, even those in capitalized format and marked before and after as must follow.
In terms of plan b4 act, I give it a long prompt expressing what we are working on, working from which point (ie: start, fixing, updating, expanding) and what the current situation is, the goals of the entire process etc. This often works well enough for the LLM to give some code, or even logical sounding ‘game plan’, but often enough it is wrong, and even gets to a point sometimes it admits it is working off of old data it was trained on and following built in protocol to not search for new updated information even if prompted so long as it can determine with a greater likelihood that it already knows the answer (from that outdated training data).
For markdown files, I think I have them setup properly. I create them in sublime text and save them with a .md and it seems to be a markdown file. I use these sometimes to pass in code, but not often, since while I watch the LLM do it’s ‘thinking’ I can see it stating clearly it doesn’t in it’s environment have the entire file, etc. In fact, I tested 4 different LLMs once with a markdown file of around 9,000 lines of code, and not a single one of them was able to use that file to give me correct guidance, because not a single one was able to give me the exact snippet from within the entire code that I need to alter, and they just made stuff up.
So, from my experience, LLMs are fine, they work well enough that they are useful to a degree, but they provide bad answers often enough that they are not ‘AGENTS’ and can not be left alone to their own devices to really complete tasks. Especially tasks, where the existing code is beyond 4-5K lines…What I typically do is give it snippets of the areas I know to be where we need to make changes or expand functionality to, and that works better since I only provide 500-1,500 lines of code.
Did you say NEARLY perfect, or PERFECT? I think key thing to understand is I’ve been clearly talking about how the Bubble AI Agent will need to have PERFECT JSON…anything that is just NEARLY perfect is not good enough, and will inevitably result in issues.
You do not need bubble to create those IDs for you through the private API calls. You can run the same basic randomized string setup and get the same type of IDs and send those into your JSON file for the bubble app. They are not created in any way that requires Bubble to actually do it. They mostly just follow a pattern, and I do not believe Bubble is doing anything to create them first and then validate the JSON against them…I think Bubble just creates them and puts them into the JSON.
If you are trying to build your own Bubble AI app builder, and the new IDs of the data types, fields, colors, etc. are the issue at this time…it will become even harder to ensure the LLM will ALWAYS use those same exact IDs perfectly in any JSON it spits out for you, which ultimately is my basic point about how Bubble AI Agent will suffer the same inherit flaws as any LLM powered Agent at this time.
It is not Bubble, its the LLM.
Would yours be better than the Google course?
I suggest not…there are really on 5-6 main points to know and they can be summed up in 5-6 short paragraphs…Prompting is not the issue…its the LLM, as they are not providing perfect responses, nor do they always follow the existing guardrails, or keeping in memory (context) previous changes or updates made.
And that is my point…do you really think Bubble is going to go into every app and tweak the guardrails in their agent for each bubble app as each bubble app grows in size? No, they wouldn’t. They have set guardrails in place and they likely will not be adaptive and evolving.
I’m glad you are doing well with AI coding and making the manual interventions needed to be successful with it…My points were specific about a Bubble AI Agent needing to contend with the known issues in LLMs, since it is not a Bubble issue, it is an LLM issue, an issue you are also experiencing with a need to redesign guardrails as your apps codebase expands.
No it is not and it is wrong. This why people are not getting consistent results. There is so much more to it.
also ChatGPT pro via ui is bad. Simple things it can do but any meaningful and it will get a mess. It is how everybody starts, and that’s ok, but not to get something of any meaningful size.
That Bubble might not be able to pull it off is yet to be seen. It needs lots of knowledge and experience to understand how to do it right. And it is not about doing something specific for each code base as you suggest. That’s another misconception. Perhaps the hardest part for Bubble is all their custom build code that probably for good reason accumulated to spaghetti with non industry standard coding patterns. Years ago, when they received millions in funding, a radical shift by building a full ai driven platform grounded in their knowledge of the NoCode market would have made them an absolute winner by now. Today it is a big challenge that is fascinating to follow!
1 Like
Relax folks, I’m making the agent for Bubble that you’re looking for.
1 Like
My friend, you are bringing a plastic knife to a gun fight…
3 Likes
Is that supposed to export a bubble app to real code?
You could use it for that, but it’s not what it’s for. It connects your AI client (e.g ChatGPT, Cursor, Claude Code) to your Bubble app to investigate real issues with real data, logs, plan new features, come up with optimisations for existing ones, etc.
The real question is, will I successfully implement agentic editing in some form on existing Bubble apps before Bubble does 
Let me give some examples.
Find all unused backend workflows and verify with server logs
Here I want to find unused backend workflows, so it does that. But wait, what about public workflows not called internally? It checks the logs to make sure it hasn’t been called recently.
This particular app has 1,300+ backend workflows, by the way. Buildprint found 273 workflows that are not called internally, but 44 of those were called externally recently.
So, we now have 229 workflows we can get rid of from this app.
Investigating error rate
In our logs we can observe a high error rate (error executing workflow/action)
Let’s find out about it:
I’ve quickly come to learn that building on Bubble manually is actually very fast. It’s planning/getting context/debugging that eats up all time. Even if the Bubble AI agent could never edit apps, that wouldn’t be much slower than AI assisted coding given Bubble is just fast to build on.
So, I’m tackling the first problem first, and will soon be turning to editing 
2 Likes
Good luck…my curiosity is not around how that’s built, it’s about how you make an LLM not hallucinate or just give false information which causes issues.
Every implementation of an LLM has same issues…I’ve seen it first hand in a training session when student used a tool that is supposed to scan the app and tell you about it. First question asked provided false information.
Where is the AI Agent?
I have a Growth plan but don’t see where to open the AI Agent. What am I missing?
The agent is only available on AI built apps for now.