Building reliable web agents with human demonstrations

Hey everyone! Super excited to share what we’ve been working on – Optexity, a platform for training browser automation agents that work reliably on any website. With our platform you can train agents by showing human demonstration of your workflows.

We fine-tuned Llama 3-8B (10x smaller and cheaper), outperforming GPT-4o & Gemini 2.0 on complex Hubspot workflows. We will be adding more functionality to train agents from RL, docs & YouTube tutorials to reliably operate software. All demos and source code can be found on our website.

With the current repository you can create human demonstrations from which models can learn to use that software. Or simply try vanilla out of the box gemini to complete tasks on a website.

Check it out: https://optexity.com

https://github.com/Optexity/ComputerGYM

We'd love to hear from devs using or planning to use some sort of RPA or browser-use or openai Operator for their tasks. If you are having hard time fixing those workflows or these models don’t work for you please discuss. Let us know what you think and star the work if you find it interesting.

We are a team of researchers from CMU, Ex-Apple and Amazon and have been building it for over a month.

Thanks !!