Today, Zeta Labs, a London-based startup founded by former Meta engineers Fryderyk Wiatrowski and Peter Albert, announced the launch of Jace, an LLM-powered AI agent that can execute in-browser actions on command.
The company also announced it has raised $2.9 million in a pre-seed round of funding, led by Y Combinator’s former head of AI Daniel Gross and former GitHub CEO Nat Friedman.
While AI agents have been in the news lately (Cognition’s Devin being the most popular one), Zeta claims its offering doesn’t need any guidance and can save users entirely from sitting in front of their computers. They just have to tell the agent what needs to be done and it will get to work.
The startup is working with some early partners and plans to use the pre-seed money to further improve the capabilities of Jace, making it more reliable and faster to handle highly complex tasks consumers and businesses may demand. Multiple other angel investors and VC firms also participated in the round, including Shawn Wang, Bartek Pucek and Mati Staniszewski, the founder of ElevenLabs.
VB Transform 2024 Registration is Open
Join enterprise leaders in San Francisco from July 9 to 11 for our flagship AI event. Connect with peers, explore the opportunities and challenges of Generative AI, and learn how to integrate AI applications into your industry. Register Now
What kind of tasks can Jace AI agent do?
Albert first envisioned the need for an AI agent when working on an ecommerce business eight years ago. He and his team had to do a lot of mundane operational work, like moving data from one source to another. Fast forward to the GPT age, when language models were mature enough, he decided to team up with fellow Meta engineer Wiatrowski and started working on Zeta Labs and its core product — Jace.
At the core, Jace is a straightforward web agent — much like ChatGPT. You go into the chatbox, interact with the bot and describe what needs to be done. Once all task instructions are provided, either through natural language or follow-up widget-like prompts, the underlying models get to work, where they create a plan, provide information and take action in the browser.
For instance, if a user says they want to book a specific hotel in Paris for a given week, Jace will search the web (like Perplexity) for information on that hotel and go a step beyond to visit the website of the hotel and make a booking, complete with payment. Albert told VentureBeat the offering adds arms and legs to text-generating AI chatbots and can do all sorts of tasks by working in a browser in the cloud, right from basic stuff like searching for flights or replying to emails to complex tasks like setting up a recruitment pipeline on LinkedIn, managing inventory and launching ad campaigns.
In one case, it was even able to build a company – complete with a business plan and registration – and find its first client to make money.
As it takes action, the user can switch the layout of the AI agent to view how it operates on the browser.
Autonomous Web Agent under the hood
To achieve these capabilities, Jace leverages a combination of models. One is a regular LLM (best available one) that handles chat-based interaction, captures required information and creates a plan of action, while the other is Zeta Labs’ proprietary web-interaction model AWA-1 (Autonomous Web Agent-1). It converts the plan into browser action, effectively handling the challenges and inconsistencies commonly found in web interfaces.
“Our core model is based on an open-source model. We put our dataset to reinforcement learning from AI feedback (RLAIF) and fine-tuned it on top of it,” Wiatrowski told VentureBeat. He explained the company used extensive simulated interactions and synthetic data to ensure the model could handle web tasks with multiple steps.
In many cases, web agents can also go into loops when handling tasks with 10 or more steps. Wiatrowski said Jace avoids that with the use of reasoning systems that verify if the plan has been executed or not.
“It’s a different cognitive architecture, where the verifier, the planner, and all those components allow for more complexity. I think now we allow for hundreds of steps,” he noted. Jace also includes guardrails to ensure the credentials provided by the user for a particular – like LinkedIn job posting – are stored in an encrypted format, similar to that of a password store.
Release and monetization in pipeline
While Jace can already handle a range of tasks, Zeta Labs has not monetized the product yet. The company is working with a few design partners to further refine the AI agent and prepare it for general release. As part of this effort, it is also working on the second iteration of the AWA model — which will be much larger and faster as well as better at handling longer, more complex tasks, especially those requiring visual work from the agent (like interacting with maps).
Notably, most of the pre-seed funding will go towards this direction, along with some hiring efforts.
Ultimately, Zeta Labs hopes it will be able to package this agent as a lucrative sidekick to consumers as well as small businesses looking to automate repetitive browser-based tasks in sectors such as recruiting, ecommerce, marketing and sales. There will be a free plan with limits on the number of messages. Once it is exhausted, users will have to pay a fixed subscription price of $45/month.
“On the business side, especially with small businesses, we see a massive demand. A great example is recruiters who want to source from LinkedIn and move data to Airtable. Currently, the process is manual. They search with binary search strings, take the data, paste it into Airtable, calculate the internal score and then use it to do matching. This entire pipeline can be automated with Jace. You just have to ask,” Wiatrowski added.
