Cerebrum AI Agent Laboratory

Create self-evolving agents that build there own tools

Comment 1

Inspiration

Cerebrum was born from a long-standing fascination with Artificial General Intelligence, inspired by science fiction novels and films. In many ways, AGI feels like the ultimate goal for our generation. We've spent years experimenting with AI in countless ways, and today, AI can perform many tasks we typically do online. Our ability to delegate responsibilities and oversee their execution through intelligent systems is simply too powerful not to harness.

The concept of “Jarvis,” a fictional AI assistant that feels naturally integrated with its environment and user, no longer feels out of reach. The tools and infrastructure are already here—now it's about building agents that can evolve, learn new skills dynamically, and operate beyond the chat interface.

What it does

Cerebrum introduces a new approach to building intelligent agents. Rather than relying on manual tool connections or research, we developed a system for dynamic tool generation. This allows agents to autonomously discover, test, and integrate tools suited to the user’s specific needs.

The Cerebrum ecosystem simplifies the full lifecycle of agent creation—from design to execution—making it more seamless, intelligent, and autonomous. It aims to move past the rigid "visual programming" model used by many tools today, proposing a more flexible and adaptive method for building AI agents.

How we built it

Many users, including myself, found limitations with tools like Bolt.new. But eventually, I realized the issue wasn’t the tool—it was the workflow. Tools have boundaries, yes, but it’s up to the developer to understand and navigate them effectively.

In my experience, success comes from maintaining your own context rather than relying entirely on chatbot memory. Clearly defined tasks, organized into grouped objectives and then broken down into subtasks, proved to be the most reliable approach. I used large context windows not for generating fluff, but for outlining high-resolution task plans that remained consistent. Modern models often hallucinate or drift mid-task, so tracking progress independently is crucial.

Challenges we ran into

This project came with its fair share of obstacles:

1 GitHub Suspension: On day two of the hackathon, my GitHub account was unexpectedly suspended. I wasn’t given a clear reason, and with limited time, I opted to create a new GitHub and a new Bolt account. Since Bolt doesn’t allow changing the GitHub linked to a Bolt profile, I worked with the Bolt team to find a workaround for submitting my project. I’m grateful for their support.

2 Trusting AI with Terminal Commands: For the first time, I trusted AI to run terminal commands for me. This was both exciting and a little terrifying. But in the end, it contributed significantly to the agent’s autonomy and performance.

Accomplishments that we're proud of

Built a fully functional prototype capable of dynamic tool generation for AI agents

Enabled agents to independently gather, test, and integrate tools based on specific tasks

Introduced context-controlled pipelines to reduce hallucination and task drift

Developed a task-driven workflow where agents can self-correct and revise with minimal human input

What we learned

I apparently lack soft skills—I speak like an AI avatar (according to a friend who kept reminding me to "blink while talking").

The platform doesn’t matter—what matters is the prompt and process. Every tool is just a vehicle; it's how you drive that counts.

What's next for Cerebrum, AI Agent Laboratory

This hackathon wasn’t just a competition—it was a launchpad. Winning (or even participating) in the world’s largest hackathon validates not only the idea, but the urgency behind it. Cerebrum’s mission is to bring self-generating intelligence to every AI agent.

We believe in a future where agents don’t just follow scripts—they build their own toolsets, adapt, and grow. This hackathon accelerated that mission, and we’re just getting started.

NOTES

Please note that Bolt project that is linked to this submission is a backup fork of another Bolt chat, this is the chat initially created the project https://bolt.new/~/sb1-y7szrnnt, , ive used Bolt backup function in order to start new Bolt chat.

I was advised to mention this information ( that ive used bolts backup to create clone of the project) as a note, but i was made aware that judges see that anyway

Built With

  • bolt
  • nextjs
  • supabase
  • vercel-ai-sdk
Share this project:

Updates