Kitesurf: Agent-first browser that runs in V8 isolates
TL;DR Highlight
Cloudflare launched Kitesurf, an agent-first browser running on Workers, in free beta
Who Should Read
Backend developers whose agents use headless browsers and who find Chromium's memory and CPU costs too high
Core Mechanics
- Chromium is overkill for agents
- Tabs, themes, extensions, and device sync don't matter to AI.
- Agents care about token count, context windows, scalability, performance, and cost.
- Per-instance memory and compute make one Chromium per agent prohibitively expensive, per the post.
- It runs entirely on Workers
- Kitesurf runs only on Cloudflare Workers and is free during beta in Browser Run.
- It is significantly more efficient in CPU and memory than Chromium for screenshots and HTML extraction.
- The excerpt gives no concrete numbers.
- It started by porting obscura with AI
- The team tried porting obscura, a Rust headless engine, to Workers using an AI agent.
- It worked poorly at first.
- It worked once they gave a solid plan and clear success criteria so the agent could loop and ask questions.
- Tests kept AI-written code under control
- Web Platform Tests (WPT) served as the agents' goalposts.
Community Reactions
- Blitz's author weighed in
- A commenter says Kitesurf is built on Blitz and that they were told Cloudflare intends to open source and upstream its patches.
- The article excerpt doesn't mention open sourcing.
- The V8 isolate guards only half the risk
- One commenter argues the danger is not code escaping but the agent reading hostile pages and misusing its normal tools.
- Prompt injection doesn't need to break out of V8, they say.
- Identity, not rendering, was the blocker
- A commenter who tried raw HTTP across the open web reports being stopped by CAPTCHA, SMS verification, and cards on file.
- They argue an isolate browser fixes none of these.
- Questions on anti-bot and fingerprinting
- One comment asks whether Kitesurf gets special bypass privileges on Cloudflare-protected sites.
- Others say it would be easy to fingerprint and block, and wish it were open source for local use.
How to Apply
- Measure your current browser task cost first
- Pick the screenshot or HTML extraction tasks your agent runs today and measure memory and CPU per Chromium instance.
- Save these numbers as the baseline for comparing against the Kitesurf beta.
- Point your existing CDP code at Browser Run
- Change only the connection endpoint of an existing Puppeteer or Playwright script and run the same job.
- Compare the resulting HTML and screenshots against Chromium.
- Cut browser-agent tool permissions to the minimum
- List the tools your agent can call after reading an untrusted page.
- Remove write and send tools from read-only tasks.
- Define tests before delegating to AI
- Before handing code to an AI agent, set the test suite and success criteria it must pass.
- Tell the agent to ask questions when unsure.
Try It
# Paste into Claude or ChatGPT (simplified version of the post's approach: auditing browser tasks for agents)
I use a headless browser with an LLM agent. Classify the task list below.
Task list: {{tasks the agent performs in the browser}}
Split each task into:
1. Light rendering is enough (screenshots, HTML extraction)
2. Needs identity (login, CAPTCHA, payments)
3. Risky if a hostile page is read and tools are then called
Also propose the minimum-privilege tool list for each task.
Next step: write 5 success-criteria tests for the category 1 tasks.Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.