The Multi-Agent Browser Problem
We hit a wall last month: spinning up multiple AI agents that each needed their own browser instance. Playwright works fine for one agent, one browser. But the moment you try to launch two agents sharing the same User Data Directory, the browser crashes. The constraint is hard: only one browser process can use a given userDataDir path at any time.
That forced us to research alternatives. We evaluated five platforms, self-hosted, managed, and everything in between, to understand the trade-offs.
The Contenders
Steel Browser: Open-source self-hosted
Steel Browser is available on GitHub. Deploy it with a single Docker command:
docker run ghcr.io/steel-dev/steel-browser
It launches on its default API and Chrome-debugging ports. Each concurrent session consumes 200-500 MB of RAM. The team behind it has production validation: thousands of free-tier users, hundreds paying. In October 2025 they doubled concurrent session limits. They also built an MCP server for agent integration.
The tradeoff: you own the infrastructure. You manage Docker, RAM pressure, uptime. You get smaller community and less polish than managed platforms.
Browserless.io: Managed with framework support
Browserless.io supports popular automation frameworks via a CDP URL endpoint. Pricing:
- $140/month Starter: 180,000 units (one unit = 30 seconds of browser time) with 40 concurrent browsers
- $350/month Scale: 500,000 units with 100 concurrent sessions
They have native hooks for popular agent frameworks like CrewAI and Browser Use.
Hyperbrowser.ai: Persistent profiles and session recording
Hyperbrowser.ai prices by proxy count:
- $99/month Basic: 100 proxies
- $299/month Premium: 1,000 proxies
Key feature: persistent sessions that retain cookies, tokens, and page context across tasks. They use rrweb (open-source) for session recording.
Anchor Browser: Fully managed humanized Chromium
Anchor Browser raised a $6M seed round in 2025. They offer fully managed Chromium instances with AI Agent-planned deterministic browser tasks. If the agent can't accomplish a task deterministically, it falls back to pure AI. Production customers include Groq, Unify, and Browser-use. They don't publish public pricing; it's enterprise-focused.
Bright Data: Enterprise-grade with CAPTCHA solving
Bright Data Scraping Browser prices by consumption:
- Pay-as-you-go: $8/GB
- Volume commitment: $499/month minimum
Unlimited concurrent sessions, 99% uptime SLA, and integrated CAPTCHA solving.
What We Learned
- The Playwright constraint is real. Browsers prohibit launching multiple instances with the same User Data Directory. Multi-agent concurrency requires separate
userDataDirpaths per agent. This is a fundamental browser isolation design, not a Playwright bug. - Self-hosted vs. managed is a deployment decision, not a feature decision. Steel gives you DIY open-source tooling. Managed cloud platforms give you hosted infrastructure with self-healing automations. Different operating models, both valid.
- Pricing models cluster three ways: Browserless and Bright Data price by compute time or bandwidth consumed. Hyperbrowser prices by proxy count. Anchor doesn't publish, positioning as enterprise-only.
For local multi-agent development, the choice comes down to whether you want to hire yourself to debug RAM pressure, or hire someone else to own that failure.
Our next move is testing a self-hosted MCP integration against our agent stack, then running a cost model across the managed platforms for production workloads. The Playwright limitation forces us to pick: separate userDataDir paths per agent (adding orchestration complexity) or move the browser layer to one of these platforms entirely.
