Operations | Monitoring | ITSM | DevOps | Cloud

Stefan sits down with Eddie, senior engineer at Checkly, to talk through how an AI agent rewrote a legacy Node.js service in Go, and what it actually took to ship that to production safely.

Learn how Checkly does synthetic monitoring with real browsers to run your critical flows on a schedule. Get alerted with screenshots and traces s when something breaks, and define the whole setup as code in your repo. Chapters.

Your AI agent can set up your monitoring, report on platform health, and open a PR to fix a failing check. All this works automatically. In this webinar, Stefan (Developer Relations at Checkly, Google Developer Expert & Playwright Ambassador) demonstrates how Monitoring as Code makes Checkly agent-friendly by design, and walks through three agentic workflows live.

Playwright 1.59 added a `cli` parameter to the `--debug` flag, and this enables your coding agent to use the built-in step-through debugger. Step through the test steps, inspect network traffic at every moment, and access the accessibility tree. The same debugger is now usable without the inspector window.

"MCP burns too many tokens" is old news. Lazy MCP tool loading and page snapshots on disk fixed the token problem. We benchmarked Playwright MCP vs the CLI, and they're roughly the same. Pick your tools based on workflow, not token anxiety.

Since Playwright 1.59, the new "trace" command makes your test traces fully accessible in the terminal. All plain text, and your agent will love it. And if you're running Playwright with Checkly, our AI agent, Rocky, already reads your traces and delivers root-cause analysis before you even start digging.

An AI agent that can't open a browser is just guessing. Stefan from Checkly shows how giving AI coding agents a real browser via Playwright enables reliable end-to-end test generation and debugging, closing the quality gap created by faster, AI-driven shipping. The session compares Playwright MCP vs the Playwright CLI for agent workflows, showing that thanks to MCP spec changes, lazy tool loading, and skills, the two are now effectively just different interfaces to the same tool, with no real token advantage either way.