Repository navigation
AI Core: Rate Limits, System Rules, and Agent Roles (User Perspective + Solutions) #2197
Replies: 2 comments 2 replies
|
@BeeZy333 Hello! Thanks for posting! I've created separate issues for your requests here:
Please check each of these and let me know if they're accurate, or if we can add any details/context. You may have examples in mind that we can post. Thanks, again! |
|
Hello, again! After review, we’re not planning to add separate Plan, Code, and Review modes. Current AI models are usually capable enough to follow those instructions without a mode switch, and we’d rather keep the chat simple. If you want the agent to work in smaller steps, you can ask it directly, for example “outline a plan first, don’t change any files yet.” We’re closing this request for now, but we appreciate you taking the time to share your ideas, and we’re happy to hear more. If you've had the chance to try any of the new models out, we would be very interested in your feedback. |
Uh oh!
There was an error while loading. Please reload this page.
Hi devs! First off, the project is top-tier. Writing and building apps right on a phone is seriously awesome—something I, as a beginner, really wanted.
I'm not a professional developer and I don't know the deep codebase—I'm just a hobbyist trying to build my own projects. However, while working with AI Core, I keep running into issues that make it hard to get things done.
To avoid posting empty complaints, I teamed up with an AI to break down how these issues could be solved at the level of your code and architecture.
Below is how I see it: I described the problems in my own words, and the AI drafted the technical part to save us all time.
On free API tiers (Gemini, Groq), the agent quickly scans through files and frequently hits HTTP 429 rate limits. The chat crashes with an error, the session breaks, and context is lost. The Retry button doesn't help much here—agents are meant to be autonomous assistants, but right now you have to sit there like a security guard and babysit it.
Code implementation (drafted with AI):
Often, when hitting an error or just needing to understand what the agent is doing, you want responses in a specific format (text-only, explained in simple terms, etc.). Right now, you have to type the exact same instructions at the beginning or end of every prompt. It would be great to write this once and know the "rules" will stick. Every modern AI agent has this now.
Code implementation (drafted with AI):
systemInstruction = Content.systemInstruction { text(savedInstructions) }
{"role": "system", "content": savedInstructions}
Right now, the agent tries to do everything at once, which makes it get confused and burn through rate limits. Having a toggle would be great: for example, switching to "Planning" mode so it just outlines a plan without touching files, or switching to "Code" once the plan is ready.
Code implementation (drafted with AI):
When the agent triggers a build via tools and it fails, hundreds of lines of Gradle system output flood the chat context. This instantly burns through the tokens-per-minute (TPM) limit and hurts the overall user experience.
Code implementation (drafted with AI):
I don't claim to know all the internals. But in a recent Telegram post, there was a comment suggesting we share new ideas in Discussions.
I'm really glad the project and its plugins are open-source—so if none of my suggestions make it into a release, I'll just make a fork, maybe.
Wishing everyone bug-free code.
Thanks!
All reactions