The latest Agent has been useful, but lately its been creating more technical debt for me than it solves.
Pros:
-
When instructed precisely on ui generation it performs well and aligns for the most part with the overall ui theme of the application.
-
Its works well as a ‘quick search’ tool. So for example if I rethink the schema and it has implications for other parts of the of the application, I can prompt it to find all affected parts and make recommendations on changes. It also provides direct links to the affected parts. Therefore, it saves me the navigational and thinking time of making the necessary adjustments.
-
It works well as a troubleshooting tool. So for example if I have a workflow that’s not working as expected, I can define the expected behaviour and ask it to go and take a look at the said workflow and tell me what is broken. So once again it saves me time in that respect. Its recommendations however are not always spot on.
-
It seems to have a websearch capabilities, which means that its recommendations are not limited to Bubble native. Its not senior in this respect because it fails to mention critical pieces of information such as webhooks that send multiple responses with different status. Familiarity with this behaviour is a plus although I foresee issues with newer users.
-
It has a good understanding of scripts which is especially useful if your are not a fan of static plugins. It also has good understanding of html and css.
Cons:
- It does hallucinate sometimes. When instructed to fix basic logic i.e. status signals for the onboarding flow e.g. To-do, Pending, Done … it can spend more credits than expected simply thinking. This may vary by app size. This is especially true when not given very specific instructions. So for example if you simply instruct it to go fix the onboarding status tags and you don’t define what’s wrong with them, it will spend a lot of time (and tokens) thinking about what’s wrong with the status tags.
- When instructed and enabled to interact direct with the application, it can create ridiculously long and inefficient conditionals. This is by far my biggest pain point.
- It doesn’t try to minimize searches by making necessary schema changes to save on WU.
- It splits up actions unnecessarily. This is especially the case when using the ‘set state’ action.
- When the Agents permission is set to Ask or Auto, it does specify what changes it is going to make but its surface level and not helpful to a developer. I want to know and make changes to the technical approach, so not so much what (because that is already defined in the prompt) but how it will technicality implement the ‘what’. This creates technical debt because I either need to spend more tokens fixing the technical inefficiencies or go ahead and do it myself. If this were to be changed, it should be by opt-in because opinions may differ by virtue of familiarity with Bubble.
- When not specifically instructed to stay within the core ui theme, it doesn’t. It creates its own styles and uses them rather than the styles widely used in the application.
- The changes the agent makes are hard to track, once again it does tell you what it has done, but it doesn’t give technical specifics. Again, this is quite problematic for me.
I think if the agent asked more questions with respect to the technical implementation it would solve a lot of my problems.
Overall my experience has been largely positive, but some fine tuning here and there would be special.
Since Bubble is going down the AI Agent route it could be good to have a dedicated Agent category on the forum.
Shout out to the Bubble Team
.