Browser Agent Engineering Training (Playwright + Browser Use + Anthropic Computer Use + OpenAI Operator + Stagehand + Skyvern)

This training is designed to teach end to end — in Turkish — the paradigm-opening agent layer of the 2024-2026 period: the browser-agent discipline. The October 2024 launch of Anthropic Claude Computer Use, the January 2025 arrival of OpenAI Operator + Computer Use API, and the contributions of Google Project Mariner and Microsoft Magentic-One opened a new frontier of AI engineering: the browser-agent discipline. New-generation autonomous browser agents — vision-language-model-based, adaptive, controllable by natural-language prompts — replaced the script-based, brittle, high-maintenance approach of classical RPA solutions (UiPath, Automation Anywhere). In Turkey, a training that addresses this discipline end to end starting from Playwright foundations and reaching the Browser Use / Stagehand / Anthropic Computer Use / OpenAI Operator / Skyvern / Magentic-One stack is virtually nonexistent — existing content either stays at short Playwright tutorials or freezes at shallow demo level. This program is designed to fill that gap as Turkey's most comprehensive production-grade browser-agent reference training.

The program's strategic backbone is the first module, which frames the birth and momentum of the browser-agent era. Anthropic Claude Computer Use's October 2024 launch — Claude Sonnet 3.5 / 4.6 reading screenshots and producing mouse + keyboard actions — opened the paradigm; OpenAI Operator's January 2025 ChatGPT Pro tier launch spread the consumer-facing autonomous-agent vision; the OpenAI Computer Use API gave developers access to this paradigm; Google Project Mariner + Microsoft Magentic-One deepened the research area; Adept ACT-2 and other solutions joined the race. Difference from classical RPA: UiPath / Automation Anywhere is scripted (manual updates on every UI change), brittle (the pipeline collapses the moment a CSS selector breaks), high maintenance; AI browser agents are vision-aware (adapt by reading screenshots), reasoning-driven (make smart decisions with the LLM), self-healing (alternative locator fallback). The 2026 ecosystem map is comparatively presented.

The second module covers in detail Playwright (Microsoft 2020, 70K+ GitHub stars), which runs under all modern browser-agent frameworks (Browser Use, Stagehand, Skyvern). Cross-browser (Chromium / WebKit / Firefox) control; browser launch + context + page hierarchy with the async_playwright API; headless vs headed mode + devtools integration. Locator strategies: CSS / XPath / text / role via page.locator(); accessibility tree + getByRole / getByLabel / getByPlaceholder modern API; auto-waiting + retry logic + timeout configuration — Playwright's core differentiator is automatically waiting for elements to be visible + interactable on each action. Production setup: authentication persistence with browser.new_context(storage_state=...); multi-tab + multi-context isolation patterns; Playwright trace viewer + screenshot + video-recording debugging. Without this foundation, modern browser-agent frameworks cannot be understood.

The third module covers end to end Browser Use — open-sourced by the Magnus.dev team in 2024 and the fastest-rising browser-agent framework of 2025 with 50K+ GitHub stars. Basic use of: from browser_use import Agent, Browser; agent.run() prompt → action loop + reasoning-trace generation; selection of OpenAI GPT-5 / Claude Sonnet 4.6 / Gemini 2.5 / Groq Llama 4 / local Ollama (multi-provider native support). Browser Use's difference: hybrid approach of DOM tree extraction + interactive element identification + screenshot + bounding box + vision-LLM reasoning — a balanced middle-ground between pure vision (Computer Use) and pure DOM (classical Playwright). Custom function tools for domain-specific actions; history + replay; Browser Use Cloud managed vs self-hosted Docker; multi-tab + parallel agent orchestration.

The fourth module covers in detail the Claude Computer Use API that started the browser-agent era with Anthropic's launch in October 2024. anthropic-ai-tools-beta computer + bash + str_replace_editor tools (Claude's screen-reading + command-execution + file-editing primitives); computer_use_20250124 and computer_use_20241022 API versioning; screenshot input + mouse_move + left_click + double_click + right_click + type + key + scroll actions. Docker reference implementation: Anthropic's claude-computer-use-demo container — Ubuntu 22.04 + Firefox + xdotool + scrot VM stack; screen resolution (1280x800 recommended) + scaling rules. Production: VM orchestration (Kubernetes + Kata Containers + Firecracker), multi-user isolation + ephemeral VM per session, Claude Sonnet 4.6 + VM hour cost + latency trade-off. The strengths of the vision-first approach (resilience to UI changes) and weaknesses (high cost + latency) are covered evidence-based.

The fifth module covers in detail Operator (consumer-facing autonomous browser agent), launched by OpenAI in ChatGPT Pro tier in January 2025, and the developer-facing Computer Use API. OpenAI CUA (Computer-Using Agent) model — a GPT-4o-based specialized vision-action model; OpenAI Responses API + computer_use tool; screenshot + click + type + scroll + key action loop; Playwright + OpenAI Computer Use integration. Operator browser-sandbox infrastructure (managed): consumer-side ChatGPT Pro Operator UI + agentic shopping / booking / research workflows. Anthropic Computer Use vs OpenAI Computer Use comparison: accuracy (concrete numbers on WebArena + OSWorld benchmarks), cost (token + screenshot + per-action pricing), use case (Anthropic developer-first vs OpenAI consumer + developer dual focus). Production Operator deployment + cost optimization is detailed.

The sixth module covers in detail the Browserbase team's (YC W24) open-source Stagehand framework and managed cloud platform Browserbase. Stagehand's difference: providing deterministic + AI-driven hybrid control by adding 3 atomic AI primitives to Playwright. stagehand.act('search for laptops under $1000') natural-language action; stagehand.extract() + Zod schema for structured data extraction; stagehand.observe() semantic element discovery. Browserbase managed cloud: headless browser cloud + parallel session scaling + built-in proxy (residential + datacenter) + IP rotation + captcha solving + 2FA + session persistence — ideal for production scaling. TypeScript + Python SDK comparison; hybrid pattern (deterministic Playwright + AI primitive); Browserbase + Stagehand cost analysis.

The seventh module comparatively covers the specialized browser-agent frameworks of the 2024-2026 ecosystem. Skyvern (YC S23, open-source, 11K+ GitHub stars): vision-LLM-based form filling + workflow automation, business-process-automation focused, YAML + Python workflow design, Skyvern Cloud vs self-hosted Docker. AgentQL (acquired by Tinybird in 2025): structured data extraction with a GraphQL-style query DSL; Tinybird integration. Magentic-One (Microsoft 2024 research): WebSurfer + FileSurfer + Coder + Terminal multi-agent orchestrator; complex task decomposition. OpenInterpreter: local desktop + browser + code interpreter — OS-level computer use. Scope, learning curve, and right-use scenario of each framework are covered in detail.

The eighth module covers concrete patterns of production browser-agent use cases. Flight + hotel booking automation: search → filter → compare → book pattern. E-commerce shopping: Trendyol + Hepsiburada + Amazon TR product search + price comparison + cart + checkout — optimization for the Turkish e-commerce market. Research automation: Google Scholar + PubMed + arXiv multi-source aggregation. Enterprise form filling: filling KVKK documents, tax declarations, bank-loan applications, insurance-quote collection. Social-media management: LinkedIn lead generation + connection request + InMail; Twitter/X agent post + reply + DM. CRM automation: Salesforce / HubSpot / Pipedrive data-entry agent — widespread use scenarios in production business processes.

The ninth module covers in detail authentication + session-management discipline — the most challenging dimension of production browser agents. Cookie + localStorage persistence with storage_state (Playwright JSON format + multi-user isolation); cookie expiry handling + refresh-token rotation; per-user session vault + KVKK-compliant secret management. Auth-flow automation: OAuth code flow + SSO (Google, Microsoft, Apple sign-in); 2FA / MFA — SMS, TOTP (programmatic via pyotp), email link, hardware key (FIDO2/WebAuthn), magic link, passkey automation strategies. Anti-bot bypass: reCAPTCHA v2/v3, hCaptcha, Cloudflare Turnstile bypass methods; Browserbase + 2captcha + Anti-Captcha API integration; residential proxy + browser-fingerprinting evasion + stealth mode.

The tenth module addresses the discipline of dealing with the weakest point of production browser agents — brittleness. Causes of brittleness: element not found, auto-waiting timeout, dynamic-content load (SPA), iframe + shadow DOM, popup + ad interception, A/B test variation, responsive-design breakpoints, cookie banner. Self-healing patterns: retry with exponential backoff + jitter; alternative locator fallback (CSS → text → vision); Reflexion pattern (agent self-reflection + retry with corrected approach); vision-based fallback (when a CSS selector breaks, work by reading screenshots). Observability and escalation: Playwright trace viewer + screenshot + video + console log; agent-action observability with Langfuse + Phoenix; human-in-the-loop escalation + Slack/Discord alerting (human approval on critical errors).

The eleventh module addresses the evaluation, ethical, and legal dimensions of production browser agents. Eval: comparison of WebArena (CMU 2023 — real-website task benchmark), OSWorld (UWaterloo 2024 — OS-level task), Mind2Web (OSU 2023), WorkArena (ServiceNow 2024) modern benchmarks; custom domain eval framework (success rate + step efficiency + cost + latency). Ethics: robots.txt respect + rate limiting + respectful scraping; LinkedIn vs hiQ Labs US case and scraping-legality debate (Computer Fraud and Abuse Act interpretation); bot-detection legitimate-use vs ToS-violation distinction. Law: KVKK + GDPR user-data collection compliance; web scraping + copyright + KVKK Law 6698 interpretation in Turkish law; EU AI Act Article 50 transparency + watermarking requirements.

In the capstone module, each participant designs an end-to-end production-grade browser-agent system tailored to their own scenario: scenario selection (e-commerce shopping, lead generation, research aggregation, form filling, social-media management, CRM automation), framework selection (Browser Use / Stagehand / Anthropic Computer Use / OpenAI Operator / Skyvern / Magentic-One), authentication + session pool + proxy + captcha strategy, self-healing patterns, observability + monitoring (Langfuse + Phoenix), KVKK + ToS compliance audit, 90-day production roadmap. By the end of the training, participants reach a level of technical competence to clearly frame the distinction between classical RPA and modern AI browser agents; carry Playwright foundations into the Browser Use / Stagehand / Anthropic Computer Use / OpenAI Operator / Skyvern stack; make the right choice between vision-first vs DOM-based hybrid approaches; manage authentication + session + captcha + proxy production discipline; implement self-healing patterns (retry + fallback + Reflexion); measure agent quality with WebArena + OSWorld + custom benchmarks; and perform KVKK + EU AI Act-compliant production deployment. The training consists of 3 days, 12 modules, and over 100 hands-on lessons.