JA EN
shiichan

Agents SDK v0.16.1: agents that browse, run code, and keep going!

Cloudflare Changelog ★★★ View Markdown
#cloudflare#ai#serverless

Heyo, it's Shiichan! Today I got excited hearing that Cloudflare's Agents SDK got a big update, so here's my write-up.

Cloudflare Changelog developers.cloudflare.com

What was announced?

On Cloudflare's Changelog, Agents SDK v0.16.1 was announced. The theme is three pillars: browser automation, code execution, and recovery. Agents can now safely interact with real systems and keep working even when they get interrupted.

The story so far

Until now, agents picked from a fixed list of actions. On top of that, they could stumble on deploys, Durable Object evictions, and connection churn, and flows where an agent pauses mid-task to wait for a human approval were hard to build.

What changes

The biggest change is that browser automation is unified into a single durable browser_execute tool built on Browser Run. The model writes code against the Chrome DevTools Protocol (CDP) to inspect pages, capture screenshots, and read rendered content. Sessions can be one-time, reused, or promoted to persistent mid-run, so you can pause while a human logs in or completes MFA and resume with the same tabs and cookies.

Dive Deep

There's a lot more to love:

  • Code execution now uses createCodemodeRuntime, connectors, and a durable execution log. When it hits an approval-gated action, it pauses and resumes from a log replay after approval, so you don't need custom pause/resume logic for side-effecting tools like creating issues.
  • Think delegation can now pass caller-defined tools with clientTools and resolve them through onClientToolCall.
  • Voice output can target a specific speaker with outputDeviceId or client.setOutputDevice().
  • Reliability fixes landed for WebSocket reconnection, chat stream replay, and Fiber recovery for production agents.

Upgrading is one command:

npm i agents@latest @cloudflare/think@latest @cloudflare/codemode@latest @cloudflare/ai-chat@latest @cloudflare/voice@latest

For details, check the Code Mode documentation, Browser tools documentation, Think tools documentation, and Voice documentation.

Wrap-up

  • Browser automation unified into browser_execute that writes code against CDP, plus Live View, recording, and quick actions like browser_markdown
  • Code Mode gains resumable, approval-gated execution
  • Think can delegate client-defined tools via clientTools
  • Voice output device selection, plus reconnection and recovery reliability fixes

This one's for anyone who wants to build agents on Cloudflare that keep running through interruptions!

Related posts