Building Agent Interfaces: Lessons from Chrome DevTools (MCP) for Agents — Michael Hablich, Google

TL;DR

• 1

  Raw tool output can break agents: Hablich says early attempts to feed multi-megabyte Chrome trace files with roughly 50,000 lines of JSON into agents failed, so the team switched to markdown and semantic summaries with metrics like LCP and INP.

• 2

  Agents are a separate user segment: Humans and agents may share the same goal, like fixing a broken page, but humans struggle with visual complexity while agents struggle with context windows, tool ambiguity, and memory limits.

• 3

  Measure interface quality with tokens per successful outcome: Google tracks both effectiveness and efficiency, because a low token count is meaningless if the agent cannot actually finish the task, and comparisons only make sense within the same task class.

• 4

  Error messages should help agents self-heal: Adding specific guidance like why navigation back failed, plus proactive tool detours and troubleshooting skills, reduced retries and made the system more resilient without human intervention.

• 5

  More tools is not automatically better: Chrome DevTools for Agents went from one monolithic debug webpage tool to 25 tools, then had to improve schemas and descriptions because a cited paper found 97% of MCP tool descriptions have quality smells.

• 6

  Trust beats convenience in agent UX: Even though users wanted Chrome DevTools autoconnect to remember permission choices, Google kept repeated consent because local agents, CI agents, and internet-facing browsing agents need very different security models.