Descriptions:
Michael Hablich, Product Manager for Chrome DevTools at Google, shares four engineering lessons from building Chrome DevTools for Agents — a purpose-built MCP server that lets AI agents debug, performance-profile, and audit web pages directly, compatible with Gemini CLI, Claude Code, Codex, OpenClaw, and any MCP-capable agent harness.
The project was motivated by a concrete failure: coding agents could generate web code but couldn’t validate it in a real browser. Early attempts to give agents raw performance trace files — 50,000-line JSON payloads multiple megabytes in size — blew through context windows entirely. The fix was semantic summarization: instead of the raw trace, the MCP server now returns structured markdown with key metrics like Largest Contentful Paint and INP, pointing the agent at the right information rather than forcing it to read the whole book. This reframe — agents as a distinct user class sharing human intent but requiring radically different interfaces — drives all four lessons.
Hablich then covers practical MCP server design tradeoffs in depth: hiding niche tools (like Chrome extension debuggers) behind command-line flags, a “slim mode” exposing only three tools to minimize token burn at the cost of reduced agent capability, CLI chaining to shift token-heavy post-processing off the model entirely, and structured error messages to reduce costly retry loops. He introduces “tokens per successful outcome” as a north star metric, arguing that even an imperfect measurement enables data-informed decisions over pure intuition.
📺 Source: AI Engineer · Published June 05, 2026
🏷️ Format: Deep Dive
