Browserbase

A working tour of Playwright's four layers, Browser, BrowserContext, Page and Locator, with code showing how each maps to a real browser process, an isolated cookie and cache scope, and a tab. Covers the auto-waiting checks that make locators more stable than raw CSS selectors.

The case for training and evaluating multimodal agents against live browsers: a single page state is simultaneously a screenshot, a DOM, an accessibility tree, a network trace and a replay, and only a live, adversarially inconsistent web supplies all of them at once.

Browserbase splits reading the web from driving it. Fetch takes a URL and returns page content as HTML or an image with no browser session, cutting the latency, cost, and infrastructure overhead of launching Chrome just to see what is on a page.

Browserbase and Prime Intellect open an RL environment for browser agents that runs on real websites rather than static datasets, handling orchestration, task execution and verification. Validated by fine-tuning Qwen3-VL-8B on filtered WebVoyager navigation tasks.

A threat model for browser agents, where injection stops being a wrong answer and becomes a wrong action. Walks through invisible payload forms, why the attack surface covers every page, ad, script and embedded document, and why containment and deterministic automation scale better than detection.

This Browserbase guide explains how CAPTCHA solvers detect, compute answers for, and submit responses to CAPTCHA challenges using model-based recognition, human-in-the-loop solving, or token-based methods. It also documents Browserbase's built-in CAPTCHA solving feature, which is enabled by default in its browser sessions and emits console events so automation code can wait until solving completes.

Browserbase traces its shift from pure vision-based computer use to letting agents write and execute code against the browser, arguing code generation outperforms clunky click/type tool calls on long-horizon tasks.

Browserbase breaks a legacy scraping stack into five swappable primitives: Fetch for static pages, Proxies/Browsers for pages that resist automation, Stagehand for shifting markup, and Verified for hard-blocking sites, with a concrete migration path.

This explainer traces Chrome DevTools Protocol from its origins as WebKit's remote debugging plumbing to its role as the JSON-over-WebSocket interface that tools like Puppeteer, Playwright, and browser agents use to control Chromium. It walks through launching Chrome with remote debugging enabled and writing a minimal WebSocket client that navigates a page and listens for lifecycle events via CDP domains like Page, Network, and Runtime.

Stagehand v3 turns AI browser automation into a driver-agnostic, multi-language interface supporting Python, Go, Java, Ruby, Rust, and more, plus a REST API for MCP and CLI callers. It also adds parallel control of many browsers at once and includes a stealth argument for running without bundling Playwright.

How Browserbase built a runtime that puts automation code next to the browser. Covers latency compounding from cross-region CDP calls, why stateless serverless assumptions fit browser work badly, and the function, deployment and invocation split that stops every deploy from being a production deploy

Why AI reshaped coding and support before marketing: those loops are verifiable and taste is not. Hightouch's fix for hallucinated logos and fonts was a coding loop that emits SVG against a per-brand context layer, and the next unlock is giving agents a browser to read live signals.

A framework for building a natural language to Playwright code generator for any site: a prompt and editable-script UI, Anthropic for generation, and a live remote browser connected over CDP so the generated script is run against the real page before you keep it.

Why embedded frames wreck automation scripts, and how Stagehand approaches them: pull the accessibility tree via Accessibility.getFullAXTree, pair it with DOM.getDocument for structure, and build a single semantic page representation an LLM can act on across frame boundaries.

Launching one Chromium container is a weekend project; running thousands is a standing system. This piece walks through the five layers between an agent and the open web, covering warm pools, isolation, an identity sites accept, session observability, and model routing, along with what owning each one costs a team.

Aaron Levie's case for the model overhang: compute compounds while procurement, security review and process do not, so the gap between what models can do and what enterprises run keeps widening. Coding went first because code is text, access is easy, and output is verifiable by tests and CI.

Browser agents rediscover every site from scratch each run and pay the exploration tax forever. Autobrowse runs a real task until the approach converges, then graduates it into a markdown skill plus deterministic glue that any later agent or teammate can load and execute.

Browserbase exposes an x402 endpoint so an agent can discover the service through the x402 Bazaar, read per-session pricing and settle in USDC per browser session without an API key or billing account. Coinbase Business handles settlement on the receiving side.