Build anything onchain: https://thirdweb.com

Frames v2 starter: https://github.com/thirdweb-example/farcaster-starter

https://warpcast.com/~/channel/thirdweb

Releasing a new architecture for @thirdweb Nebula!

It’s now cheaper, faster, and more accurate.

Over the past couple of months, we’ve tested 5+ different agent architectures, each a step forward in speed and accuracy.

What’s changed? 🧵

Context:

Nebula typically uses 3+ tools to perform basic onchain actions -- understanding chains, researching contracts/tokens, then executing

Each tool call is a model inference (slow + costly), and outputs are interdependent. That means we can’t just merge them into a single tool

In the new architecture, Nebula generates and executes code snippets to complete tasks.

This is where the magic happens:

It thinks in code.

The snippets may use one or multiple tools, reducing the number of inference round-trips

The result? A huge improvement from the previous architecture

Most onchain actions now complete with just 1-2 inference calls

As foundational models get better at coding, Nebula will become smarter and more capable of handling complex tasks

We're constantly experimenting with crypto AI agents -  this is just the start!

More exciting updates coming soon. :)

so the LLM now returns code based on snippets that it then runs in one vs doing 3 different calls to another LLM tool that then does the call?

We're constantly experimenting with crypto AI agents -  this is just the start!

More exciting updates coming soon. :)

https://blog.thirdweb.com/nebula-v0-0-14-code-agent-real-time-streaming/