20.3Detailed Analysis: Key Competitors
Playwright CLI + Test Agents
Strengths:
- Same battle-tested engine as Playwright proper — auto-wait, actionability, tracing
- Disk-first snapshots: ~4x token savings over MCP (Microsoft's own benchmark)
- First-party planner/generator/healer agents, scaffolded per agent runtime (
init-agents --loop=claude|vscode|codex|opencode) browser.bind(): agent, test runner, and MCP share one browser instance- Traces (with network capture) as a built-in audit trail
Weaknesses:
- Needs filesystem access — sandboxed agents must fall back to MCP
- Disk round-trips can be slower in wall-clock time than MCP streaming
- Agent definitions are version-coupled (regenerate on upgrade)
Best for: Coding agents with filesystem access — which is the standard setup in 2026. The default choice.
Playwright MCP
Strengths:
- Richest in-context page understanding via accessibility trees
- Works in sandboxed environments with no filesystem
- Published to the official MCP Registry every release
Weaknesses:
- ~4x the token cost of the CLI path for typical tasks
- Context bloat on long sessions
- Microsoft itself now points coding agents to the CLI first
Best for: Sandboxed agents, MCP-native IDE integrations, quick exploratory sessions.
Stagehand (Browserbase)
Strengths:
- The standout agentic SDK of 2025-26 (~22.8k stars; large weekly npm downloads across the Browserbase ecosystem)
- Intent-level API —
act("click the submit button")— no selector authoring at all - Backed by Browserbase's hosted browser infrastructure
Weaknesses:
- TypeScript-first; ties you to its abstraction layer
- Intent resolution consumes model calls — costs scale with UI complexity
- Younger engine than Playwright's
Best for: Product teams building agent features that need browser control, TypeScript shops, hosted-browser workflows.
browser-use
Strengths:
- Most popular OSS AI-browser-agent framework (~81k stars)
- Python-native — fits ML/data teams
- Vision + DOM hybrid understanding for nonstandard UIs
Weaknesses:
- Vision calls add latency and cost per interaction
- Weaker actionability guarantees than Playwright-engine tools
- Python-only
Best for: Python teams, complex/custom UIs where selectors are impractical, research-grade agent work.
Vibium
Covered in depth in the case study. Summary: architecturally instructive (BiDi-native, server-side actionability, skills-first — it argued the CLI-skills thesis before Microsoft adopted it), shipped V1 in June 2026, but with no production adoption signals yet. Track it; don't bet a framework on it today.
Selenium 4.x
Strengths:
- Universal browser support, massive ecosystem, enterprise-grade maturity
- Low-level WebDriver BiDi today; Selenium 5 (unreleased) will bring high-level BiDi APIs
- Language support: Java, Python, JS, C#, Ruby, Kotlin
Weaknesses:
- Not designed for AI agents — no skill or MCP interface
- No auto-wait/actionability (manual explicit waits)
- Heavier setup than modern alternatives
Best for: Enterprise teams with existing Selenium investment; pair it with an agent-native tool for new AI-driven work rather than replacing it wholesale.