A founder who keeps arriving early
Furqan Rydhan has a habit of being early to the platform shift. He was a founding team member of AppLovin, the mobile advertising company that grew into a business worth roughly $180 billion, and later co-founded thirdweb, the platform that tried to make building web3 apps as straightforward as building web apps. He also co-founded Founders Inc, the venture studio built around investing in itself and building things with people it likes. The throughline is not a single technology — it is a taste for infrastructure, the unglamorous layer other builders stand on.
His current chapter is agents. Nebula, his AI workspace, is a Slack-like surface where every channel holds an agent that can execute real work across the tools teams already use — creating and editing a Google Slides deck end to end, generating images, retrying on failure until the job lands, and turning one-off work into repeatable “recipes” with scheduled triggers. In a 2026 podcast walkthrough, Rydhan showed Nebula creating content on a schedule, and described it as a step toward the one-person business powered by agents. Coverage in October 2026 noted Nebula bringing outside agents — including Claude and Codex — into the same shared workspace, charging per seat while metering its own agents separately.
The problem underneath the agent demo
Anyone who has watched an AI agent use a browser has seen the same failure mode: the agent confidently clicks something that is not there, fills a field that does not exist, or reports a price it never actually read. Browser automation built on hand-written selectors breaks the moment a site changes, and agents that guess at page structure hallucinate with confidence. The demo looks magical until it has to touch money, credentials or an irreversible action — and then nobody wants to let it.
fastbrowse, the open-source project from Rydhan's Agent Labs team, is built as the answer to that specific distrust. It does not start from a screenshot or a DOM dump. It indexes every button, link and field a page actually has, and its System One model, Jev, picks the next step from that list. It cannot click something that is not on the page, and there are no selectors to maintain. If it is unsure, it falls back to a standard approach rather than inventing an action. The design choice is philosophical as much as technical: an agent should only ever do what the page provably allows.
Answers you can check, spending you can bound
Under the hood, the choice model is Jev, built by Typesafe, which picks from the indexed controls rather than free-generating actions — the mechanism that makes “cannot click what is not on the page” literally true instead of aspirational. And because the agent runs through a skill file that coding tools like Claude Code and Cursor can adopt, a developer can hand a browser task to the agent they already work with, rather than learning a new harness for every site.
The second half of fastbrowse's bet is verifiability. Every claim in a finished answer cites a verbatim quote from the page, with a link that scrolls to it, and passing a schema returns validated data rather than a plausible paragraph. In the project's recorded demo, fastbrowse shops a live store — adding a Raspberry Pi Pico 2 W and an in-stock sensor to a basket, filling in checkout without paying — in 37 steps, about four minutes, and $0.29 in model calls. The number matters because it is measured, not promised: a browser agent whose cost and behavior can be observed is one a team can actually budget for.
The third half is control. Credentials, irreversible actions and spending stay under code, not under the model. Submit, pay, delete and send all stop until you allow them. A secret is typed only on the origin it was declared for, and models only ever see its name. Step, time and dollar limits are enforced in code. It is the same instinct that runs through Nebula — agents doing real work — with the guardrails made explicit: autonomy, but with a leash held by software rather than hope.
Infrastructure, again
fastbrowse ships the way infrastructure ships: a CLI, a Python library, Node and Go SDKs that need no Python, an MCP server, and a skill for coding agents, under an MIT license. It runs against a managed cloud browser by default, or local Chrome, or any Chromium browser or Electron app already running. The free plan needs no card and carries $2 of runs a month; paid plans start at $29 a month, with per-run pricing that never drops below half a cent.
For Rydhan, it is a familiar move in a new market: after advertising infrastructure, after web3 developer infrastructure, and after an agent workspace, the browser itself becomes the layer. Agents are about to touch every website on earth. Someone has to build the part that makes sure they only click what is really there, cite what they really read, and stop before they spend what they should not. fastbrowse is Agent Labs' argument that the trustworthy browser agent — open, cited, and bounded by code — is the layer the whole agent economy will need.
The playbook
- Agents should do real work, not demos — From Nebula's shared agent workspace to fastbrowse's browser layer, Rydhan's projects aim at agents executing across tools teams already use — not chatbots that describe work.
- Verifiability is the feature — Every claim in a finished fastbrowse answer cites a verbatim quote from the page, with a link that scrolls to it — the answer can be checked, not just believed.
- Code holds the keys — Submit, pay, delete and send all stop until code allows them, and secrets are typed only on the origin they were declared for — autonomy with a leash.
- Build the layer, not just the app — fastbrowse ships as a CLI, Python library, Node and Go SDKs, an MCP server and a coding-agent skill — infrastructure other builders can stand on.