Post

Joan León
@nucliweb.net

I'm testing the @debugbear.com MCP with Claude Code, and my first impression is that combining DebugBear's great tools with the power of AI Agents is awesome for creating executive and detailed technical reports... #WebPerf #PerfTools #AI #WebPerfAI

Reid Marlow
@reidmarlow.com

The report angle is where MCP makes the most sense to me: let the agent gather evidence, but keep the final call tied to traces, budgets, and URLs. What would you use as the acceptance test?

Joan León
@nucliweb.net

Simple test: every conclusion needs an `evidence` field pointing to real data (Lighthouse, RUM, waterfall), not generated text. If the agent can't cite it, it doesn't make the report. Testing it with any specific workflow?

0 reposts1 like

Reid Marlow
@reidmarlow.com

That's the right constraint. I'd make the evidence field a URL + timestamp + raw metric, not a summary. Otherwise the agent just moves the hallucination into the citation box. What are you testing first?