I read about Nyne this morning, a father-son founded startup that just raised funding to solve what they’re calling the “human context gap” for AI agents. The pitch is that AI agents operate in a vacuum, disconnected from what humans are actually doing, and Nyne wants to build a layer that feeds human context into those agents so they can act more intelligently. It is a real problem. But the solution already exists, and it is not a new middleware layer.

The context gap is real, the fix is not where you’d expect

Every time you ask ChatGPT to help you with something on a webpage, you have to copy the text, paste it into the chat, explain what page you were on, describe the layout, mention what you were trying to do. That is the context gap. The AI has no idea what you’re looking at, what you clicked before, what tab you have open, or what form you’re halfway through filling out. You are the manual bridge between your browser and your AI, and frankly, it is exhausting.

Nyne’s bet is that this bridge should be a platform, a context layer that captures what humans are doing across their tools and pipes it into AI agents that run elsewhere. And I get why that’s fundable, because the problem is obvious and painful, and building infrastructure for AI agents is the kind of thing VCs love to write checks for right now.

But I keep coming back to a simpler observation: the richest source of human context is the browser, and if your AI agent already lives inside the browser, you do not need a separate context layer at all.

What “seeing what you see” actually means

When an AI agent runs as a Chrome extension in your side panel, it has access to the DOM of the page you are on, which means it can read every heading, every table cell, every form field, every email thread, every product listing, every paragraph of the article you are halfway through reading. It does not need you to copy-paste or explain. It does not need a middleware platform to reconstruct your context from event logs and screenshots.

So when you’re staring at a spreadsheet in Google Sheets and you ask the agent to summarize column B, it just reads column B. When you’re on a LinkedIn profile and you want a draft outreach message, the agent already knows the person’s title, company, recent posts. When you’re deep in a GA4 report trying to figure out why your direct traffic spiked, the agent can walk through the same nested menus you’re clicking through, because it is right there with you, looking at the same screen.

This is not some futuristic capability. It is how browser extensions have always worked, except now the extension has an LLM behind it instead of a static script.

Middleware solves the wrong problem

Nyne’s approach assumes that AI agents live outside the browser and need context piped in. And for some agents, that is true. Coding agents in your terminal, background automation runners, agents operating on server-side data — those genuinely lack human context because they’re not in the loop where humans work. A context layer makes sense there.

But most knowledge work happens in a browser. Email, docs, spreadsheets, CRMs, project management tools, research, shopping, social media, admin consoles. If the AI agent is already present in that environment, building a separate system to capture and relay context is like installing a security camera in a room you’re already standing in. You don’t need the camera. You have eyes.

The overhead of a context middleware is not just technical, either. It is a privacy question. Capturing human context across tools means routing sensitive data — emails, documents, credentials screens — through yet another third-party system. Browser-native agents like dassi avoid this entirely because the context never leaves your browser. The page data goes directly from the DOM to the LLM call, and nothing gets stored or relayed through an intermediary, which matters a hell of a lot more than most people realize when they’re handing over access to their work environment.

The agents that win will be the ones already there

I suspect the context gap will keep getting funding because it sounds like infrastructure, and infrastructure plays are catnip for venture capital. But the pattern I keep seeing is that the most useful AI agents are not the ones with the most sophisticated context pipelines. They’re the ones that skip the pipeline entirely by being present where the work happens.

Nyne might build something genuinely useful for headless agents and backend automation. And the founding story is compelling. But for the 90% of knowledge work that happens inside a browser tab, the context your AI agent needs is already rendered on screen, sitting right there in the DOM, waiting for something smart enough to read it.