Skip to main content

Lab Notebook · Agentic browser

A browser that was an agent platform, not a window — built before Arc and Dia turned that way

In June 2025 we built a desktop browser whose core idea was that the browser itself should be an agent platform, not just a viewport onto web apps. It ran as an MVP with autonomous agents and a virtual office system. The honest edge: it was a direction proven, not a product finished — and the market was about to get crowded.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 20 June 2025
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 20 Jun 2025
Living document. Reviewed 20 Jun 2025
Entry date
20 June 2025
Category
Agents
Lead over the world
~1 year ahead
Access
🔓 Public
ResearchPartly provenDownload PDFPDF · 732 KB

We built a browser on the premise that the browser itself should be an agent — a place agents live and act — right as the industry started heading the same way.

VOAS
virtual office agent system, working
Electron
desktop shell hosting the agents
1 thesis
'the browser is an agent platform' — validated

Honest evaluation

Partly proven

The 'browser as agent platform' thesis ran and was real; the product-grade safety, reliability, and permission model it would need were not built or proven here.

What would prove or disprove it further

What would extend the proof: harden the permission and safety model to the point an agent can act inside a real logged-in session without a human watching every step, and measure a failure or abuse rate under adversarial use. Absent that, "the thesis is architecturally sound" stands, but "this is safe enough to ship" remains untested.

The evidence — full reasoning behind the verdict

Verdict: partly-proven. The core architectural claim — that a browser can be built as an agent-hosting platform rather than a passive viewport — was demonstrated true: autonomous agents ran as first-class citizens inside the Electron/FastAPI shell, with a working virtual office agent system coordinating them. What wasn't proven is the harder, adjacent claim a real product would need: that this can be done safely and reliably enough to trust with someone's logged-in browsing life.

The mechanism behind the "ran" half of the verdict is concrete — a working MVP with a WebSocket-linked backend, agents visibly acting inside the browsing context. The mechanism behind the unproven half is just as concrete: the report says outright that the permission surface for an autonomous agent acting inside an authenticated session was prototyped, not solved, and that the polish and hardening a shipping browser needs were never built. Those aren't hedges added after the fact — they're the honest boundary of what an MVP can show.

By mid-2025 we had already built browser extensions that let an AI read pages, fill forms, and navigate (ChromeBot, and the chat-extension line before it). The next question was bigger: what if the browser itself were the agent platform, not a window you bolt an extension onto?

What it was

vocabotics Browser was an AI-native desktop browser. Its organising idea was that the browser should be a place agents live and act, with autonomous agents and a virtual office agent system (VOAS) handling tasks on your behalf — rather than a passive viewport onto other people's web apps.

What we built

A working MVP as an Electron desktop shell with a Python/FastAPI backend and a WebSocket layer, carrying:

  • Autonomous agents operating inside the browsing context.
  • A virtual office agent system for organising task automation.
  • A glassmorphism UI over the whole thing.
Research

The measured claim is modest and honest: the thesis ran. You could watch agents act inside the browser as first-class citizens of the environment. That was the point — to see whether "the browser is an agent platform" was a real architecture or just a nice phrase. It was real.

What we learned — including the honest negatives

  • An MVP proves a direction, not a product. The thesis worked; the reliability, security hardening, and polish that a browser actually needs to be trusted with your logged-in life were not built here, and we did not pretend they were.
  • Agents in the browser is a safety problem. An autonomous agent acting inside your authenticated browsing session is powerful and dangerous in equal measure. We prototyped the permission surface; we did not solve it fully, and that gap is honest.
  • Timing beats stubbornness. Shortly after, the wider market — Arc, Dia, and others — started heading exactly this way. The right move was to bank the validated direction and not spend a year fighting for a product slot that was about to get crowded.

Where it went / status

Archived as a validated MVP. The browser-as-agent-host idea folded into the lab's broader agent platform work rather than shipping as a standalone browser. What it bought us was evidence — that we saw where interfaces were heading before the market did — and a clean example of the discipline of stopping a proven experiment before it becomes a sunk cost.

What is still open — kept visible

The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.

  • It was an MVP: the thesis ran, the polish and reliability of a shipping browser did not.
  • Browser-hosted autonomous agents raise real safety and permission questions we prototyped but did not fully solve here.
  • The market moved fast; validating the direction was worth more than fighting for a crowded product slot.

Where this connects

Sources

  1. vocabotics project audit — vocabotics Browser (Electron/FastAPI), Jun 2025vocabotics internal project history · as of June 2025

    We use cookies.