What is Browser Use and how does it compare to Playwright?

Browser Use is an AI agent library where the LLM plans and executes all browser actions autonomously from a natural language goal. Playwright is a deterministic scripting framework where you write explicit step-by-step code. Browser Use achieves 89.1% on the WebVoyager benchmark and adapts to UI changes automatically but costs $0.02-$0.30 per task. Playwright achieves ~98% reliability at millisecond speeds with zero marginal cost, but requires manual script updates when UI changes.

What is Stagehand and why does it exist?

Stagehand is a TypeScript SDK built by Browserbase on top of Playwright. It adds three AI primitives (act, extract, observe) while keeping full access to Playwright's deterministic capabilities. This means you can write a workflow with Playwright selectors for the stable parts and Stagehand AI calls for the variable parts — capturing the reliability and cost benefits of Playwright where it works, and the flexibility of AI automation where it doesn't.

What is Browserbase and how does it relate to these tools?

Browserbase is a cloud service providing managed headless Chrome browsers with persistent sessions, anti-detection, and parallel execution. It is the backend for Stagehand (they built it) and also works with Playwright and Puppeteer. It handles the browser infrastructure — you handle the automation logic. Hermes Agent integrates Browserbase as one of its browser backends for tasks requiring cloud-hosted, persistent browser sessions.

Why does Hermes Agent support 4 different browser backends?

Different tasks have genuinely different requirements. Browserbase is best for cloud-hosted persistent sessions with anti-detection. Browser Use is best for complex autonomous navigation. Chrome CDP is best for tasks needing low-level browser control. Local Chromium is best for completely private/air-gapped workloads. Rather than picking one approach and forcing everything through it, Hermes allows per-task backend configuration to match the tool to the job.

How much does AI browser automation actually cost per task?

Browser Use: $0.02-$0.30 per task (5-20 LLM steps, vision tokens for screenshots). Stagehand: $0.002-$0.02 per individual AI action (cheaper per step, no vision by default). Playwright/Puppeteer: zero marginal cost for the automation itself — only the server running the browser (which you already have). At Hermes OS on the Pilot plan, a daily competitor monitoring task running on Browser Use typically costs $0.05-0.15 per run, or $1.50-$4.50/month.

AI agent browser automation tools compared: Browser Use, Stagehand, Playwright, and Puppeteer

The tool landscape in 2026

Four tools dominate AI agent browser automation in 2026. Browser Use is the pure autonomous agent approach — the LLM observes the page and decides everything. Stagehand bridges AI and deterministic automation via three clean primitives. Playwright is the industry standard for traditional scripted automation, now increasingly used as a foundation layer by the AI tools above it. Puppeteer is the older Chrome-specific alternative to Playwright, still widely deployed but increasingly superseded.

The core tradeoff runs along two axes: reliability vs. flexibility, and cost vs. capability. A pure AI agent (Browser Use) handles any page without brittle selectors but costs more per run and is slower. A pure deterministic script (Playwright) is near-instant and free at volume but breaks on every UI change. The 2026 production pattern for most teams is a hybrid: AI handles the variable, unstructured parts; Playwright handles the high-volume, stable parts.

Hermes Agent supports all four approaches. Browserbase (cloud headless browsers), Browser Use (autonomous agent loop), Chrome CDP (raw DevTools Protocol), and local Chromium are all available as backends depending on the task requirements. The choice of backend is per-task configuration — not a global setting.

Browser Use: the autonomous agent approach

Browser Use is an open-source Python library (MIT license, 50,000+ GitHub stars) that wraps a full autonomous agent loop around browser control. The LLM observes the current page via screenshots and DOM extraction, decides on the next action, executes it, and repeats until the goal is achieved. No explicit step-by-step scripts required — you specify the goal in natural language and the agent figures out how to navigate to it.

Benchmark performance: 89.1% success rate on the WebVoyager benchmark, the standard evaluation suite for web navigation agents. Task completion rate in production: 72-78% depending on model (GPT-4.1 or Claude 4.6). Performance breakdown by task type — simple action: 2-5 seconds; form fill: 10-30 seconds; data extraction: 5-15 seconds. Cost: $0.02-$0.30 per task depending on complexity (5-20 LLM steps per task, each consuming vision tokens). Script breakage rate: under 5% — the AI adapts to UI changes without manual selector updates.

Where it breaks down: it is the slowest option (full agent planning loop), the most expensive per task, and the hardest to debug (requires reading traces rather than looking at a script). For tasks with inconsistent target sites or open-ended goals — 'find the cheapest available flight' — it is the right call. For tasks that run at high volume against stable sites, the cost and latency disadvantages compound significantly. Hermes Agent uses Browser Use as its default autonomous browsing backend for complex research tasks where the exact navigation path is unknown.

Stagehand: the hybrid AI/deterministic approach

Stagehand is an AI web automation SDK built on Playwright by Browserbase (TypeScript/JavaScript, MIT license, 10,000+ GitHub stars). It exposes three AI primitives: act() for natural language actions, extract() for structured data with Zod schema validation, and observe() to identify page elements. Version 3.0 (released in 2026) communicates via Chrome DevTools Protocol directly and is 44% faster than v2.0.

Benchmark performance: approximately 75% task completion rate on WebVoyager. Performance breakdown — simple action: 1-3 seconds; form fill: 5-15 seconds; data extraction: 2-8 seconds. Cost: $0.002-$0.02 per action (one order of magnitude cheaper than Browser Use for individual steps). Script breakage rate: under 5% over 30 days.

The distinctive value: Stagehand lets you mix AI and deterministic steps in the same workflow. Login with explicit selectors (reliable, free), then use act('click the export button') for the part that changes monthly. This hybrid model captures most of the reliability benefit of pure AI automation while avoiding the full cost of running every step through an LLM. Cloud hosting is available via Browserbase at $0.01/minute of browser time. Stagehand is the recommended approach for TypeScript teams building production workflows with a mix of stable and variable UI elements.

Playwright: the deterministic baseline

Playwright (Microsoft, Apache 2.0, free and open source) is the industry standard for scripted browser automation in 2026. Cross-browser support (Chromium, Firefox, WebKit), auto-waiting for elements, network interception, tracing, and parallel execution across Browser Contexts. Multi-language support: JavaScript, TypeScript, Python, Java, C#.

Benchmark performance: approximately 98% task completion rate on known navigation paths — the highest reliability of any option when the UI is stable and the script is current. Performance: simple actions under 100ms, form fills under 500ms, data extraction under 200ms. Cost at volume: zero (open source, no per-request pricing). Maintenance burden: 15-25% of scripts break over 30 days when target sites update — the core operational cost of the deterministic approach.

Playwright is the right tool when: you are running the same page structure thousands of times per day (cost and speed matter), the target site is an internal tool or API with stable structure, or you need pixel-perfect compliance for a financial or regulated workflow. It is the wrong tool for: sites that change frequently, competitor monitoring tasks where you do not control the target, or any task where the exact UI path is unknown. The 2026 Playwright AI ecosystem at currents.dev describes the direction: 'Playwright is now turning testing workflows into programmable building blocks that agents can drive.' Playwright-as-foundation, AI-as-planner.

Puppeteer: the Chrome-native legacy choice

Puppeteer (Google, MIT license) is the older Chrome-specific alternative to Playwright, using the Chrome DevTools Protocol directly. It remains widely deployed — many production scrapers were built on Puppeteer in 2020-2023 and have not been migrated. For new projects in 2026, Playwright is the cleaner default.

The functional difference: Puppeteer has deeper, more direct access to Chromium's internals — CDP sessions, security settings, performance profiling, service worker interception. For Chrome-specific automation that needs low-level DevTools access, Puppeteer remains the right call. For general cross-browser automation or any task where the Chromium-only limitation is a problem, Playwright is superior — better auto-waiting, explicit Firefox/WebKit support, and a more consistent API across browser engines.

In multi-agent AI systems, Puppeteer-as-CDP-backend is a pattern Hermes Agent explicitly supports: raw Chrome DevTools Protocol access is available as a backend option for tasks that need low-level browser control alongside the higher-level AI agent planning layer. This is relevant for stealth scenarios (anti-bot bypass, specific fingerprint requirements) where the CDP interface gives you precise control over browser identity characteristics.

Which tool to use for which job

High-volume, stable-site automation (internal tools, structured APIs, CI/CD testing): Playwright. Near-zero cost, near-100% reliability, fastest execution. Maintenance is manageable when you control the target or can alert on breakage quickly.

Variable public-site research, competitor monitoring, open-ended navigation: Browser Use (via Hermes). 89.1% WebVoyager benchmark, adapts to UI changes automatically, handles multi-step goals without scripting. Budget $0.05-0.30 per research task at Haiku+Sonnet rates.

Production hybrid workflows (stable login + variable content extraction): Stagehand. Mix deterministic selectors for known steps with act()/extract() for variable ones. 44% faster in v3.0, Zod-typed output, cost between the two extremes.

Chrome-specific surveillance/fingerprint tasks or raw CDP integration: Puppeteer via Hermes Chrome CDP backend. Deep DevTools access where Playwright's cross-browser abstraction is a constraint rather than an asset.

The Hermes Agent approach on Hermes OS: Browserbase handles cloud browser sessions (persistent, anti-detection, parallel); Browser Use handles autonomous goal execution; Chrome CDP handles low-level tasks; local Chromium is available for completely offline/private workloads. All configured per-task, not globally. Vision paste capabilities allow the agent to analyse screenshots directly without DOM parsing — useful for sites with complex JavaScript rendering or canvas-based interfaces.