Relay built AI agents that could actually operate web interfaces — click, type, navigate, extract — not just summarize pages. Their demo videos showed agents booking flights, filling expense reports, navigating SaaS dashboards. The hard part was never the LLM reasoning; it was the execution layer. Handling dynamic DOMs, shadow roots, iframe hell, CAPTCHAs, rate limits, and the thousand edge cases that make web automation brittle. They solved enough of it to attract Google's attention.
Chrome already has the runtime. CDP (Chrome DevTools Protocol) gives programmatic control over every tab, network request, and DOM mutation. What's been missing is a reliable semantic layer — something that understands "click the submit button on the third step of this checkout flow" without hardcoded selectors. Relay's team spent two years building exactly that semantic layer. Now they're inside the organization that controls the browser engine.
This matters for anyone building AI workflows today. The current pattern — LangChain/LlamaIndex agents calling Playwright or Puppeteer via CDP — works but fractures at scale. You're managing browser pools, handling session persistence, fighting memory leaks, and debugging flaky selectors at 2 AM. If Chrome ships a native "agent mode" with built-in semantic understanding, the entire browser automation stack compresses into a single API call.
const result = await chrome.agent.run({
task: "Find the cheapest round-trip SFO to NYC next Tuesday, book window seat",
constraints: { maxPrice: 400, airline: ["United", "Delta"] }
});
That's the direction this points. Google has the distribution (3B+ Chrome users), the rendering engine, the V8 team, and now the Relay crew who know how to make LLMs drive browsers reliably. They don't need to expose a standalone product — they can embed it in Chrome Extensions API, in DevTools, in the new AI-assisted browsing features they've been teasing.
For developers, the practical takeaway: don't over-invest in custom browser automation frameworks right now. The abstraction layer is about to move down the stack. Playwright and Puppeteer won't disappear, but the "agentic" layer on top — the part that translates intent to actions — is likely to become a platform capability rather than a library you maintain.
Also worth noting: Relay raised $5M from Sequoia and others in 2023. The exit wasn't a blockbuster, but the team landed in a position to influence how billions of users interact with the web. That's the real outcome for AI infrastructure talent right now — not necessarily a unicorn exit, but placement where your work becomes the default path.
Curious whether Firefox or Safari respond, or if this becomes another Chrome-only standard that the web has to adopt. Either way, browser automation just got a lot more interesting.
Next Training image models on a laptop actually works now →
All Replies （0） #
No replies yet — be the first!