#21 · AI Agent Applications
Top Browser-Use & Computer-Use Agents
What is a browser-use or computer-use agent?
A browser-use or computer-use agent is an AI agent that operates a browser or full computer the way a human would — moving a cursor, clicking buttons, typing text, scrolling through pages, filling out forms, switching tabs, navigating multiple applications, and completing multi-step workflows across interfaces that weren't built for AI access. The category sits in a different architectural family from API-based agents: rather than calling clean structured APIs, these agents perceive the screen via screenshots (and increasingly via structured DOM representations or accessibility trees), reason about what they see, and take actions through simulated mouse and keyboard inputs. The category effectively started with Anthropic's October 2024 Computer Use research preview, then expanded rapidly through 2025–26 as every major lab shipped browser or computer-use capabilities and a wave of dedicated browser-agent products emerged. By mid-2026, the landscape splits into four architectural families: *dedicated agentic browsers* (Perplexity Comet, OpenAI Atlas) where the browser itself is the agent; *OS-level agents* (Anthropic's Claude Cowork on macOS, Claude in Chrome extension) that operate at the desktop level; *cloud-hosted virtual browsers* (OpenAI Operator, Google Project Mariner) running in vendor-managed sandboxes; and *open-source agent harnesses* (Browser Use, Stagehand) for developers building their own browser automation.
Why browser and computer-use agents matter in enterprise AI.
The economic argument is concrete: vast amounts of enterprise back-office work happens through web UIs and desktop applications that don't have clean APIs — government portals, ERP systems built decades ago, supplier order systems, e-commerce admin interfaces, insurance claim systems, tax filing portals. Robotic Process Automation (RPA) addressed some of this work but required brittle scripts that broke when UIs changed. Browser and computer-use agents address the same workloads but with adaptability — vision-based agents auto-adapt to small UI redesigns (button moves, color changes) in ways that classic RPA can't. The 2025–26 capability inflection is real: Anthropic's Claude Sonnet OSWorld score jumped from under 15% in late 2024 to 72.5% by February 2026 (after Anthropic's Vercept acquisition pushed vision-based perception forward), approaching human baseline performance on the standard benchmark. Per-task cost dropped from $0.50–1.50 in 2024 to roughly $0.05–0.15 in 2026, making SME deployment economically viable for the first time.
What to evaluate.
Browser and computer-use agent selection should consider: (1) capability scope — browser-only vs. full computer control (Claude Computer Use and Cowork support full OS control; Operator and Comet are browser-only); (2) deployment model — managed cloud sandbox vs. self-hosted sandbox vs. extension on user's actual browser; (3) security model — credential isolation, action confirmation thresholds, audit logs, prompt-injection defenses; (4) benchmark performance on OSWorld and WebArena for relevant task types; (5) integration with broader agent workflows; (6) enterprise compliance posture. Prompt injection remains a fundamental vulnerability — Anthropic reported that unmitigated agents fall for 24% of prompt injection attacks, with defenses cutting that rate by more than half but never to zero. Production deployment requires sandboxing, human approval for sensitive actions, and explicit checkpoints before agents access financial accounts or credentials. The list below ranks ten browser and computer-use agents most defensible for enterprise consideration.
Frontier computer-use agent with desktop and browser control
Anthropic's computer-use capability — exposed via the Claude API in beta and as the production Claude Cowork desktop agent (launched March 2026) — currently leads the category on OSWorld performance after the February 2026 Vercept acquisition pushed Claude Sonnet 4.6 to a 72.5% score. The architecture is fundamentally different from cloud-sandboxed alternatives: Claude operates on the user's actual desktop, controlling native apps and browsers through the same interfaces a human would use. Available through Anthropic's API, Amazon Bedrock, Google Vertex AI, and direct via Claude Cowork on macOS. Best for organizations needing full desktop and OS-level automation, complex multi-application workflows spanning browser and native apps, agentic coding workflows where computer control is integrated with code agents, and Claude-native development stacks. Strengths include category-leading OSWorld performance (72.5%), full OS-level control (not just browser), broad cloud availability (Anthropic API, Bedrock, Vertex), integration with the broader Claude Agent SDK and MCP ecosystem, and Zero Data Retention arrangements for organizations needing it. Trade-offs are higher per-task cost than narrower browser-only agents, beta status for parts of the API, complexity of safe deployment on user desktops, and the security implications of giving an agent full computer control.
Cloud-hosted virtual browser agent integrated with ChatGPT
OpenAI's Operator launched in early 2025 as a standalone product running OpenAI's Computer-Using Agent (CUA) in a cloud-hosted virtual browser; the standalone Operator product was sunset in mid-2025 and absorbed into ChatGPT Agent Mode, making browser agent capabilities a feature within the broader ChatGPT product. The architecture runs the agent on OpenAI's cloud infrastructure in a managed virtual browser, with the agent pausing and handing control back to the user for sensitive actions like logins. Operator is currently trailing Claude on OSWorld (38.1% vs. 72.5%) but maintains the broad ChatGPT ecosystem integration that drives consumer and enterprise adoption. Best for ChatGPT-standardized organizations, web-task automation where the cloud-sandboxed model fits, applications integrating with DoorDash/Instacart/OpenTable/Uber (which have direct partnerships), and ChatGPT Plus/Pro subscribers wanting browser agent capabilities as part of the broader subscription. Strengths include integration with the broader ChatGPT product, partner integrations with major service providers, managed cloud sandbox reducing security complexity, and consistent ChatGPT developer ecosystem. Trade-offs are no persistent memory across sessions (agent starts fresh each time), no local file or calendar access (browser-only), reliance on OpenAI's cloud infrastructure for the browser session, and benchmark performance trailing Claude.
Dedicated agentic browser for consumer and prosumer use
Perplexity Comet is a dedicated agentic browser — the browser itself is the agent, with the address bar functioning as both URL input and prompt box. The browser handles tasks like "research flight prices to Dubai and summarize options" by opening tabs, reading sites, and compiling results autonomously. Following a Q1 2026 cross-platform rollout (iOS in March 2026 after Android in November 2025), Comet now spans macOS, Windows, iPad, Android, and iOS. The platform became broadly accessible (previously $200/month Max-only) in October 2025. Amazon's January 2026 lawsuit challenging Comet's automated shopping is the first major legal test of agentic browsing. Best for prosumer and consumer research workflows, applications combining browsing and research synthesis, organizations valuing Perplexity's broader research ecosystem, and users wanting an agent-first browser as their primary interface. Strengths include category-leading agentic-browser cross-platform availability, integrated research synthesis with browsing, broad consumer adoption, and Perplexity's existing research and citation capabilities. Trade-offs are dedicated-browser commitment (users switch browsers, not just install an extension), pending legal questions around automated shopping, and less suited for enterprise governance than developer-facing alternatives.
OpenAI's dedicated agentic browser with ChatGPT integration
OpenAI Atlas, launched in October 2025, is OpenAI's dedicated agentic browser — ChatGPT integrated into every tab, with an Agent Mode that autonomously browses the web and completes tasks. Atlas combines context-aware sidebar interactions, memory of user preferences across sessions, and direct ChatGPT ecosystem integration (custom GPTs, conversation history). In March 2026, OpenAI announced Atlas would merge with ChatGPT and Codex into a single desktop "superapp," making the standalone Atlas roadmap somewhat fluid. The OpenAI Computer-Using Agent powering Atlas achieved 87% success rate on WebVoyager and 58.1% on WebArena in internal benchmarks. Best for ChatGPT subscribers wanting deep ChatGPT integration in browsing, applications leveraging custom GPTs and ChatGPT memory, organizations standardized on OpenAI's broader stack, and users wanting a dedicated agentic browser within the OpenAI ecosystem. Strengths include deep ChatGPT integration, memory across sessions (unlike Operator), strong WebVoyager benchmark performance, custom GPT and conversation history integration, and active OpenAI development. Trade-offs are currently Mac-only (Windows/iOS/Android in development), strategic uncertainty around the planned superapp merger, and resource consumption higher than lightweight alternatives.
Browser extension bringing Claude into existing Chrome installations
Claude for Chrome takes a fundamentally different approach from Comet and Atlas — rather than a dedicated browser, it's a Chrome extension bringing Claude's capabilities into the user's existing browser. Launched in August 2025 as a limited preview for Max subscribers, expanded to Pro/Team/Enterprise in December 2025. The extension can take actions on websites, fill forms, and integrate with Claude Code for debugging workflows. The March 2026 Quick Mode update makes the extension approximately 3× faster by bypassing standard tool-use protocols in favor of a compact command language. Claude for Chrome puts significant emphasis on security with site-level permission controls and action confirmations for sensitive operations. Best for organizations wanting agentic browsing without switching browsers, users staying on Chrome but wanting agent capabilities, applications integrating with Claude Code for development workflows, and security-conscious deployments valuing fine-grained permission controls. Strengths include extension-based deployment (no browser switching), strong security model with site-level permissions, integration with Claude Code workflows, Quick Mode for fast browsing, and access to multiple Claude model tiers (Sonnet 4.5, Opus 4.6). Trade-offs are Chrome-only deployment, less seamless than dedicated agentic browsers, and limited to browser context (not full OS).
Google's browser automation research with Chrome integration
Project Mariner is Google's research prototype for browser automation, currently available to Google AI Ultra subscribers and increasingly integrated into Chrome through the Gemini 3 side panel. Mariner handles tasks like finding job listings, hiring service providers, and ordering groceries by interacting with websites autonomously. The platform identifies its web requests with the Google-Agent user agent (since March 2026), giving website owners visibility into agent-driven traffic. Chrome's January 2026 Auto Browse feature extended Mariner-style capabilities to Chrome Premium subscribers. Best for Google AI Ultra subscribers, Google Cloud–standardized organizations, browsing workflows benefiting from Google's broader ecosystem (Gmail, Calendar, Maps integration), and applications leveraging Gemini 3's multimodal capabilities. Strengths include deep Chrome and Google Workspace integration, multimodal capabilities through Gemini, Google's broader cloud infrastructure, and transparent agent identification via Google-Agent user agent. Trade-offs are Google subscription required for full capabilities, less open development than Anthropic and OpenAI alternatives, and the research-prototype framing means production-readiness is still evolving.
Open-source browser agent framework with broad model support
Browser Use is the leading open-source browser agent framework — hitting 81,200+ GitHub stars by March 2026 and becoming the fastest-growing open-source browser agent framework. The framework provides flexibility in model choice (works with OpenAI, Anthropic, local models), self-hosted deployment, and direct integration with the developer's actual browser profile (existing logins work without re-authentication). Best for developers building custom browser automation, self-hosted browser agent deployments, applications needing flexibility in model choice, cost-sensitive deployments avoiding per-call vendor pricing, and use cases where browser-profile integration (existing logins, cookies) matters. Strengths include open-source license, broad model compatibility, very active development with 81K+ GitHub stars, can use real browser profiles with existing logins, and strong community of contributors. Trade-offs are higher technical complexity than managed alternatives, requires developer engagement for deployment, and less polished out-of-the-box experience than commercial products.
Managed cloud browser infrastructure for AI agents
Browserbase provides managed, cloud-hosted headless browsers purpose-built for AI agents — solving the operational complexity of scaling browser automation in production. The platform handles browser lifecycle management, fingerprinting, anti-bot evasion, session persistence, and the infrastructure work that makes browser agents practical at scale. Increasingly used as the underlying infrastructure for both Browser Use-based applications and proprietary browser agent products. Best for developers building production browser agents needing managed browser infrastructure, applications requiring high-scale browser automation, organizations wanting to avoid the operational complexity of running headless browsers, and platforms providing browser agent capabilities to their own users. Strengths include category-leading managed browser infrastructure, mature production deployment patterns, strong anti-detection capabilities, broad SDK and API support, and clear positioning in the infrastructure layer. Trade-offs are infrastructure-only positioning (not an end-user product), per-session pricing that requires evaluation for high-volume use cases, and dependency on Browserbase for the underlying browser execution.
Microsoft's agentic browser capabilities in Edge
Microsoft's Edge Copilot Mode, redesigned and rolled out through April–June 2026, transforms Edge into a Copilot-first browser with agentic capabilities. The strategic positioning is that Edge becomes a Copilot-native browser rather than a traditional browser with AI added on, with deep Microsoft 365 integration and access to the broader Copilot ecosystem. Best for Microsoft 365 enterprise customers, organizations standardized on Edge for corporate use, Copilot-first deployments, and users wanting agentic browsing integrated with Microsoft enterprise tools. Strengths include Microsoft enterprise ecosystem integration, Edge's existing enterprise deployment footprint, Microsoft 365 integration, and Copilot subscription leverage. Trade-offs are Edge browser commitment (users switch browsers), Microsoft ecosystem lock-in, and less mature than dedicated browser agent products from AI-native companies.
Production-grade browser automation framework
Stagehand, an open-source framework from Browserbase, provides production-grade primitives for browser automation built on top of Playwright — combining AI-driven actions ("act," "extract," "observe") with traditional Playwright reliability. The framework targets developers who want AI capabilities for the genuinely dynamic parts of browser workflows while keeping stable, high-volume flows scripted in Playwright. Best for production browser automation combining AI agents with traditional scripting, developer teams transitioning from Playwright to AI-augmented automation, applications requiring reliable production browser workflows, and organizations using Browserbase for browser infrastructure. Strengths include production-oriented design (built on Playwright reliability), open-source framework, clean integration with Browserbase infrastructure, and pragmatic hybrid AI-plus-scripting approach. Trade-offs are developer-focused (not for non-technical users), requires Playwright familiarity, and narrower than full agent frameworks for non-browser workflows.