Michael Hablich is Google’s product manager for Chrome DevTools and Puppeteer, helping coding agents inspect, test, and repair working websites instead of merely generating code. His signature project, Chrome DevTools MCP, connects AI development tools directly to Chrome’s debugging and browser-automation capabilities.
Based near Vienna, Hablich previously worked as a software developer, tester, agile coach, program manager, and engineering manager. His early writing explored product failures and customer value and measuring the impact of process improvements. He also contributed to the 2021 MDN Web Testing report, a collaboration involving Google, Microsoft, and Mozilla that examined browser compatibility and practical obstacles to web testing.
In September 2025, Hablich and Mathias Bynens introduced Chrome DevTools MCP, an open-source server that lets coding agents inspect console errors, network activity, layouts, user interactions, and browser-performance traces. Its official GitHub repository supports compatible agent environments without tying developers to one coding assistant. An early design delegated debugging to another agent inside the server; Hablich’s team replaced that extra layer of unpredictability with focused browser tools that leave reasoning to the calling agent.
- Agents need purpose-built interfaces. Hablich treats agents as a distinct user group: humans navigate visual interfaces, while models need explicit schemas, clear tool descriptions, and bounded responses. His work on agent-tool discoverability addresses the trade-off between providing useful capabilities and overwhelming the model’s context window.
- Tokens per successful outcome. His agent-efficiency metric combines task completion with token consumption, tool calls, and elapsed time. Comparisons should stay within similar workflows: extracting page content and diagnosing a broken responsive layout require fundamentally different amounts of work.
- Semantic performance summaries. Raw browser traces can overwhelm a model before it identifies the problem. Hablich favors concise, actionable findings about Largest Contentful Paint and Interaction to Next Paint, supplemented by pagination, scoped toolsets, and command-line processing outside the model’s context. Clear error messages and troubleshooting playbooks help agents recover without human intervention.
- Browser-agent trust boundaries. His browser-security framework distinguishes supervised local browsing, isolated continuous-integration environments, and internet-connected agents. Explicit consent protects authenticated browser sessions; separate profiles, container isolation, domain restrictions, and prompt-injection defenses address progressively broader risks.
Hablich also applies his testing background to evaluating LLM-powered applications, combining human review, automated raters, benchmarks, and real-user experiments to determine whether AI products actually work for their intended users.
