Experimental AI Agents/WebLLM playground

This is experimental, run at your own risk!
There is a remote possibility that running this example might even crash your computer - take appropriate precautions before hitting the "run" button below.
The results are meant as examples only, and as is common with large language models you might get surprising results, depending on which model you select.
Those results are not meant to be used aynwhere except in this experiment.
Note also that this page downloads large model files to your computer, possibly several gigabytes, when you run the agents below.

This page is an experiment with client-side Large Language Models, meant to help you understand the basics of how AI agents work. It runs 3 agents locally in your browser, using WebLLM.

An input text, found below, is fed to the first agent, and other agents get the output of the previous one as their input.

The first run may take some time as the language model downloads to your browser. After that, the example agents will take a few seconds to a few minutes to run, depending on the selected model and your computer's power. Try a smaller model initially, as indicated by the "size" value in the list of models below.

To run this, your browser needs to support WebGPU. You'll get an error message if that's not the case, in the "output" section below. Recent versions of Chrome work, see caniuse.com or use the WebGPU report.

After starting the agents with the below button, look at the "output" section at the end of this page to see what's happening and get the results.

Agent Definitions

In this section, you can tweak the agent definitions as much as you want. Use the "run" button below to re-execute them after modifying their parameters. Starting without changes is recommended so that you get some initial results to help you understand how this all works.

The code that runs these agents is quite simple, and all agents run the same code. It is only the textual behavior definitions below that make them perform different actions.

Input

Many models support other languages besides English, try it!.

Output

You can select different models to compare their performance and output.

Status: Initialized.