Ex-Meta engineers launch Jace to handle browser tasks autonomously

Most AI tools are great at giving you advice, but when it comes to actually booking the flight or updating your spreadsheet, you are still the one doing the clicking.

Two former Meta engineers want to change that. London-based startup Zeta Labs just emerged from stealth with Jace, an autonomous AI agent designed to take over your browser and complete multi-step tasks from start to finish.

Alongside the product launch, Zeta Labs announced a $2.9 million pre-seed round led by former GitHub CEO Nat Friedman and Daniel Gross, with backing from ElevenLabs founder Mati Staniszewski and other prominent angels.

Giving chatbots hands and feet

Co-founder Peter Albert came up with the concept years ago while running an e-commerce business, tired of copying and pasting data between messy web dashboards. Once modern LLMs arrived, he teamed up with fellow Meta engineer Fryderyk Wiatrowski to build a proper autonomous assistant.

Interacting with Jace feels familiar at first. You type your goal into a chat window, answer any follow-up questions, and the agent takes it from there. Unlike typical copilots that need constant babysitting, Jace is designed to let you close the tab and walk away.

Under the hood, Jace runs on a two-tier cognitive architecture. A standard frontier LLM handles conversation, extracts requirements, and sketches out an action plan. Then, Zeta Labs' proprietary web interaction model, dubbed AWA-1, executes those steps inside a cloud-hosted browser.

Albert claims the agent can handle everything from travel reservations and LinkedIn job posting workflows to inventory management. In one internal test, Jace reportedly drafted a business plan, registered an entity, and acquired its first paying customer. For tasks requiring login credentials, sensitive information is kept in an encrypted vault.

Beating the 10-step loop

Anyone who has played with web agents knows their biggest flaw: after about ten clicks, they almost always get stuck in a bizarre infinite loop.

Wiatrowski says Jace avoids this through a dedicated reasoning system that constantly verifies whether each sub-step actually worked. Built on top of an open-source model fine-tuned with synthetic data and RLAIF, the setup can reportedly handle workflows with hundreds of consecutive browser actions.

The product is currently being tested with design partners and has not yet been monetized. The team is already working on AWA-2, which aims to incorporate visual understanding for tricky web components like interactive maps.

Why it matters

We are finally watching the industry shift from text generation to actual autonomous execution.

For knowledge workers and small business operators, an agent that can reliably handle mundane web chores is the holy grail. Think of spending five minutes setting up a recruitment pipeline or syncing messy merchant portals instead of losing half a day. But right now, we think caution is warranted: handing an early-stage model full access to your corporate accounts and credit cards carries obvious operational risks.

From a technical perspective, Zeta Labs' approach of splitting high-level reasoning from low-level browser interaction is smart. Dedicated small models trained specifically on DOM trees and UI mechanics are far cheaper and faster than brute-forcing full-sized LLMs on every button click.

Still, the open web is notoriously hostile to automation. Websites change their layouts constantly, bot-detection systems are aggressive, and dynamic interfaces break easily. Proving that Jace can survive the messy real-world web over hundreds of steps will be the real test.