Browserbase

Browserbase chased browser latency through bigger machines and thinner virtualization before finding the real cause: sequential CDP commands multiply every round trip, so physical distance sets the ceiling. Going multi-region took a month and more than halved task time.

Browserbase explains why launching Chromium in an agent's existing code sandbox works locally but struggles in production. It separates the agent sandbox from Chromium's own process sandbox and covers compute pressure, security and observability.

Browserbase's Director leaves experiment status. Describe a web task in plain language, watch it run in a real browser, then export the Stagehand code the successful run produced and deploy it in parallel on Browserbase. Sensitive steps hand off to a human or to 1Password.

Browserbase Functions now support Secrets. Create a credential via API, SDK or CLI, attach it to specific Functions, and read it at runtime through context.secrets so keys stay out of source code.

Patching Chrome from JavaScript could not make automated browsers behave like real ones, so Browserbase forked Chromium. The post traces how one navigator property is assembled across threading, network and binding layers, and the cost of rebasing patches every six weeks.

Compares computer use, where agents read screenshots and click pixel coordinates, with browser use through the DOM and accessibility tree. Argues most web tasks need only a browser and outlines when a full desktop is warranted.

Browserbase explains why it abandoned rrweb DOM-mutation replay for browser sessions. Nested iframes, canvas and shadow DOM went unrendered, and replayed pages pulled live data, producing recordings that looked correct while misrepresenting the run.

The browser, not the model, is usually where agent automation breaks. This lays out four thresholds — unattended execution, concurrency, crash recovery, and sites that block automated traffic — and what changes when the browser becomes a CDP-addressable remote service instead of a window.

Seven fixes for the ways browser automations break, drawn from support tickets: reuse the preconfigured default context instead of spawning a new one, persist logins through the Contexts API, drive PDF handling with CDP download behavior, and route proxies deliberately rather than by default.

An open catalog of 100+ reusable browser playbooks, plus an open-source CLI that loads them on demand. Each skill is a markdown file with the steps, selectors, undocumented endpoints and fallbacks for one site, so agents stop paying the discovery tax on every run.

How a browser infrastructure team approached a 5x speedup: build benchmarks from observed customer workloads, add OpenTelemetry tracing and CDP traffic analysis, and keep running the benchmark while shipping. The argument is that point-in-time benchmark snapshots become vanity metrics.

Roughly 4,000 browser hours across 200-plus runs benchmarking computer-use models on OnlineMind2Web and WebVoyager, with every model tested through its public API under identical step, timeout, and stealth constraints. The write-up also covers how to parallelize evaluation from hours to minutes.

This explainer defines the Model Context Protocol as an open standard (introduced by Anthropic in late 2024) that lets AI models call tools, read resources, and use prompts through a single interface instead of custom per-model integrations. It walks through the client-server JSON-RPC exchange, STDIO vs Streamable HTTP transports, and uses Browserbase's hosted browser MCP server (navigate, act, observe, extract) as a concrete example of giving a model real web-browsing ability.

Why an automated browser lands back on the login page every run — the profile on disk is the only reason your everyday browser remembers you — and how to fix it: save cookies for simple auth, or the full profile (local storage, IndexedDB, service workers) and re-attach it to later sessions.

Pairing Stagehand's natural-language browser actions with MongoDB Atlas to turn unstructured web content into an AI-ready dataset. Covers page.extract and page.act, agent.execute for multi-step workflows, and why parsing accessibility trees beats selector-based scripts on live sites.

Browserbase explains Web Bot Auth, a standard where agents sign requests to prove identity, why header and IP checks fail as bot traffic overtakes human traffic, and how identity differs from access decisions.

Browserbase opened up the recordings behind its own dashboard. Two API endpoints list a session's tabs and return an HLS playlist for any of them, so agent replays can be embedded in QA, support or ops surfaces. Recordings persist for 31 days.

Stagehand drops its Playwright dependency and talks straight to the Chrome DevTools Protocol, with a modular driver layer that also runs on Puppeteer or Bun. The team reports a 44% average speedup on iframe and shadow-root interactions, plus a rewritten context builder.

Browserbase added a search endpoint powered by Exa and tuned for navigational queries, returning ranked URLs with titles in JSON, Markdown or HTML. It targets the search then fetch then browse pattern that agents currently fake by proxying a search engine.

Browserbase joined Okta's Cross App Access ecosystem, swapping static API keys for dynamic, scoped, revocable tokens issued through user identity, giving enterprises audit trails and policy control over browser agents.

Browserbase and HUD describe how to structure browser RL tasks: prompt, environment and outcome grader, with QA agents reviewing traces. It covers pitfalls such as live-web variability, graders fooled by page edits, and agents finding public answer keys.

Browserbase runs a single Slack-resident agent that writes PRs, queries Snowflake and drives browsers, loading skills and permission scopes per task instead of shipping a bot per workflow. Covers the ephemeral sandbox, credential brokering and skill system in full.

Twenty years of CAPTCHA generations — distorted text, harder text, image grids — each eventually beaten by machines. The argument: with agents running real workflows, verification is moving from what a browser can do to who it is, via signed agent identity like Web Bot Auth.

Browserbase drops the stealth framing for browser agents and introduces signed agent identity built on Web Bot Auth, with Cloudflare, Fingerprint and 1Password as partners. The argument: consistency across sessions, not per-check evasion, is what bot defenses can actually reason about.