Here are exciting problems we’re working on to push what’s possible with AI and knowledge work.
Automating the company
The one thing more important to us than building Polar is building the company that can build Polar. We constantly think about acceleration instead of velocity: how do you design systems that scale agents to automate as much of the company as possible?
For example, forking Chromium normally takes tens of millions of dollars and a team of experts. We did it with zero experience as 3 people with coding agents. We created a framework to have a team of agents autonomously build a Windows version of the app in a little over a week.
This insane leverage doesn’t have to only apply to engineers. We use Polar to build Polar. The rapid progress of agents enables a completely new way of building companies in 2026 and beyond.
How do you build a lean team where each individual has 1000x leverage with agents? How do you design agent systems that can self-manage and self-verify to automate every role from engineering to GTM? How do you make the entire company legible to agents to build a self-improving company?
Agent-human interface
The capabilities for browser agents already exist, but they haven’t taken off. A big reason is the interface: working with agents still isn’t intuitive, and nobody has really innovated here yet. ChatGPT and Claude weren’t built for delegating 10 hours of browser work. Whoever cracks this might own the pattern a billion people use, the way the GUI or the feed did.
Our users hand off tasks that take 10,000+ LLM calls across 100s of subagents for 10+ hours. Some have workflows that trigger every day or week. Tasks appear anywhere: one-off asks, long runs while browsing, tabs delegated to different agents or taken back mid-run.
So, what are the right interaction patterns between humans, agents, and the internet? How do we design a browser a billion knowledge workers love using every day? How do we make Polar easy to use while showing users how powerful Polar can be?
Self-improving agent harness
Every week, Polar takes millions of actions across research, sales, recruiting, ops, and marketing tasks. Manually looking through agent trajectories to debug or come up with creative harness changes is a slow, but necessary process.
Can you design a high signal agent to automatically triage agent trajectories, analyze rollouts, and extract insights? Can the agent propose changes given the codebase/user data/evals and see feedback to self-improve the self-improving harness?
Compounding memory graphs
Knowledge workers spend 8 hours a day in the browser. There is so much context in the tabs they use every day, the Polar tasks they run, and data in their logged in apps.
How can Polar learn from all of this information to have the most context in every task or workflow? If the user is a part of a bigger team, then how can you create a company knowledge graph?
Browser agent evals
Code is easily verifiable and repeatable, so very friendly to RL. But a lot of knowledge work is neither verifiable (“best” is hard to define) nor repeatable (can’t “undo” a LinkedIn DM). And you can’t reproduce most trajectories since rollouts are on a user’s logged in data.
How do you design evals for browser agents given these constraints? Do you use browser agents to construct the eval state on domains you care about? Do you build synthetic websites that are usually too simplistic? Do you find comparable sites?
The blank slate problem
Most people have never used AI agents before. There are 50M+ coders, but 1B+ knowledge workers. The most popular knowledge work agent products might have several million users, or <0.3% of all non-coders.
If you’re at an AI startup, then you probably know terms like MCP servers, skills, tool calls, tokens, browser agents, subagents, and more. But a lot of people who fall in love with Polar have never heard of these and only know what ChatGPT is.
How do you bring these powerful capabilities to the masses in an easy to understand way? How do you reduce the time to the “holy shit moment”? How do you design the product to make it easier to learn the skill ceiling and what other users do?
Extremely long-running browser agents
We’ve seen Polar crush large scale, long-running tasks due to how we designed the browser agent. But…
- How do we give better ways for Polar to orchestrate organizations of agents while ensuring smooth agent to agent communication across layers?
- How do you design browser agents that can run for days, weeks, and even months?
- How can we optimize the tools we give Polar to interact with dynamic and difficult web UIs across one billion websites?
Owning the model layer
Since we are model agnostic, we mix and match frontier models into one powerful knowledge work agent based on their comparative advantages. And our users don’t care about what models we use as long as Polar delivers results.
But with enough data, can we train our own specialized models to be #1 on the Pareto frontier of knowledge work? How should we be structuring data now to best prepare for that long term future?
Most tasks that our users use Polar for do not require AGI level intelligence. How can we optimize more for speed?
Future products
We are laser focused on the AI browser as our initial core product to build for what will grow the fastest today to start a users-and-data flywheel. But as we grow, we have many ideas on expansion products:
Next browser action prediction
Browser agents are slow right now, but eventually they’ll be able to add a Google Calendar event in one second. Until then, can we predict the next browser action a user will take given their tabs, computer actions, and Polar tasks? How can we make Polar proactive, knowing what you need to do using the infinite context it has?
Computer use agents
Computer-use agents aren’t highly parallelizable on one machine yet, so that’s why we began with browser use (which is still 70%+ of knowledge work). But the next natural step after nailing the AI browser is to extend its capabilities to the computer.
Cloud agents
One of the biggest requests we get is to be able to use Polar on the go. Cloud browser agents sound attractive, but 2FA/SSO/passkeys are increasingly device-bound (you can’t do fingerprints in the cloud) and many sites ban data-center IPs. Tasks need seamless authentication, so for now, agents must run on user devices to do long running tasks.
But there are ways around this with some tradeoffs. And because Polar is a Chromium fork, we have low level control and can have Polar running in a cloud computer to enable some cool stuff. After nailing the AI browser, we want to extend its capabilities to the cloud.
API product
Many users ask for an API to run Polar’s agent from their own systems. We’re focused on the browser today, but this is also a door into RPA. That’s a market headed past $100B that runs on brittle bots that break whenever a UI changes and take consultants months to deploy.
An agent that can use any website like a human is the obvious successor. As agents get cheap and fast, they’ll eat enterprise automation. And a Polar API is the infrastructure it can run on.
But why is solving these problems important?
The Trillion Dollar Prize
Cursor/Claude Code changed how coders work, but where’s the product that lets a non-coder type for 30 seconds, and a thousand agents complete hundreds of hours of knowledge work end-to-end?
It doesn’t exist yet. And the prize is enormous: 1B+ knowledge workers earn tens of trillions a year in wages — and agents price against the wage, not the software budget. Yet the leading products have several million users at most, or <0.3% market share.
The products that win here change how a billion humans work forever and become generational companies.
Join Us
If any of these problems interest you, email us at hiring@polarbrowser.com with the 2–3 most extraordinary things you’ve done!
A few more reasons to join:
- We’re playing in one of the largest markets in the world, while the window is open. And we can win it.
- Everyone operates like a founder! We intend to stay extremely lean so each person operates with 1000x leverage by designing systems that scale agents.
- Everyone does engineering, product, design, and GTM. You will go deep on every part of the company, own a bunch of problems, and operate as your own autonomous unit. If you want to start a company someday, then this is the closest thing to already doing it.
- And we do have a strong team! The three of us come from MIT, YC, Jane Street, Citadel, Modal, Apple, and Perplexity. We also have olympiad medals, Team USA international medals, and a chess national master between us.
- We’re the best place in the world to push the frontier of browser agents. Every week, Polar takes millions of real actions on real, logged-in work across sales, recruiting, ops, and research. And we’re #1 on every major web agent benchmark, ahead of OpenAI, Anthropic, and everyone else.