Today, ZetaLabsa London-based startup founded by former Meta engineers Fryderyk Wiatrowski and Peter Albert, announced the launch of Jace, an LLM-based AI agent that may perform actions within the browser on command.
The company also announced that it raised $2.9 million in a pre-seed funding round led by Daniel Gross, former head of AI at Y Combinator, and Nat Friedman, former CEO of GitHub.
While AI agents have been within the news these days (essentially the most famous being Devin from Cognition), Zeta claims that its offering requires no instruction and saves users from having to sit down in front of the pc. All they must do is tell the agent what to do and it gets to work.
The startup is working with some early partners and plans to make use of the pre-seed money to further improve Jace's capabilities, making it more reliable and faster to handle highly complex tasks that buyers and businesses may ask. Several other angel investors and VC firms also participated within the round, including Shawn Wang, Bartek Pucek and Mati Staniszewski, the founding father of ElevenLabs.
What sort of tasks can the Jace AI agent perform?
Albert first realized the necessity for an AI agent eight years ago while working on an e-commerce company. He and his team needed to do lots of mundane operational work, like moving data from one source to a different. In the era of GPT, when language models were mature enough, he decided to team up with fellow meta-engineer Wiatrowski and commenced working on Zeta Labs and its core product – Jace.
At its core, Jace is an easy web agent – much like ChatGPT. You go into the chatbox, interact with the bot, and describe what must be done. Once all of the task instructions have been provided either in natural language or through widget-like follow-up prompts, the underlying models go to work, making a plan, providing information, and taking motion within the browser.
For example, if a user says they need to book a particular hotel in Paris for a particular week, Jace will search the net (like Perplexity) for details about that hotel and go a step further by visiting the hotel's website and making a booking and payment. Albert told VentureBeat that the offering gives more power to text-generating AI chatbots and may handle all styles of tasks by working in a browser within the cloud, from easy things like finding flights or answering emails to complex tasks like organising a recruitment pipeline on LinkedIn, managing inventory and launching promoting campaigns.
In one case, the corporate even managed to establish a business (including a marketing strategy and registration) and find the primary customer from whom it made money.
During the motion, the user can change the layout of the AI agent to see how it really works within the browser.
Autonomous web agent under the hood
To achieve these capabilities, Jace leverages a mix of models. One is a daily LLM (one of the best available model) that handles chat-based interactions, gathers required information and creates an motion plan, while the opposite is Zeta Labs' proprietary AWA-1 (Autonomous Web Agent-1) web interaction model. It transforms the plan right into a browser motion and effectively handles the challenges and inconsistencies commonly encountered with web interfaces.
“Our core model relies on an open source model. We subjected our dataset to reinforcement learning from AI feedback (RLAIF) and tuned it,” Wiatrowski told VentureBeat. He explained that the corporate used extensive simulated interactions and artificial data to make sure the model could handle multi-step web tasks.
In many cases, web agents also get into loops when processing tasks with 10 or more steps. According to Wiatrowski, Jace avoids this through the use of reasoning systems that check whether the plan was executed or not.
“It's a distinct cognitive architecture where the validator, the planner, all of those components allow for greater complexity. I feel we are able to do a whole bunch of steps now,” he noted. Jace also includes safeguards to be certain that the credentials provided by the user for a particular job – corresponding to a job posting on LinkedIn – are stored in an encrypted format, much like that of a password store.
Publication and monetization in preparation
Although Jace can already handle a variety of tasks, Zeta Labs has not yet monetized the product. The company is working with some design partners to further refine the AI agent and prepare it for general release. As a part of that effort, additionally it is working on the second iteration of the AWA model – which shall be much larger and faster, and will even be higher capable of handle longer, more complex tasks, especially those who require visual work from the agent (like interacting with maps).
It is noteworthy that nearly all of the pre-funding in addition to some recruitment efforts will go on this direction.
Ultimately, Zeta Labs hopes to supply this agent as a lucrative sidekick for consumers and small businesses trying to automate repetitive browser-based tasks in areas like recruiting, e-commerce, marketing and sales. There shall be a free plan with message limit. Once that's maxed out, users may have to pay a set subscription price of $45/month.
“On the enterprise side, especially with small businesses, we're seeing an enormous demand. A very good example is recruiters who need to pull data from LinkedIn and push it into Airtable. Currently, the method is manual. They search using binary search strings, take the information, put it into Airtable, calculate the interior rating, after which use it for matching. This entire pipeline will be automated with Jace. All they must do is ask,” Wiatrowski added.
