Remote Agents, or: How I Fixed Our Page's Performance Without Really Doing Anything
summary
Built an agent pipeline that pulls DataDog RUM browser profiler telemetry, finds hot functions via profile aggregation, spins up a local dev server to reproduce, applies code edits with an LLM implementation agent, then re-profiles and gates changes with tests and an LLM judge. A 20% self-time improvement threshold triggers draft PRs and human review before codeowners. Running April–May yields about a 25% reduction in P95 INP; regressions appear later, showing scope and cadence limits.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:34Front-end performance basics
- 4:05Automated performance needs
- 6:11Static analysis limits
- 8:38Real user monitoring
- 9:22Verifiability in agents
- 11:10Pipeline research phase
- 12:56Implementation and testing
- 14:39Review and oversight
- 15:43Pattern generalization
- 16:06Results achieved
- 18:52Challenges and limits
- 20:17Key takeaways
related talks
34mAI Agents for Performance: Ship Faster, Pay Less
23mFrom Blind Spots to Merged PRs: Continuous Agentic Performance Optimization
18mAgents Don't Do Standups: Building the Post-Engineer Engineering Org
33mYou are the main thread: Multiplying your productivity with agents
30mAgents Building Agents
45mAI agents in practice - beyond the prototype
17mSoftware Development Agents: What Works and What Doesn't
46mLeveraging AI and Automation in DevOps