Modernizing Our Oldest, Scariest Code with AI
summary
LaunchDarkly uses agentic development patterns to modernize a 66,000-line React targeting system under a six-week timeline with a $10,000 inference budget and strict no-customer-disruption rules. The team makes the repo agent-ready, uses Meticulous for visual regression checks, and breaks work into 22→34 phases, generating ~39,000 new lines and spending ~$7k on inference. Autonomous agents accelerate generation but require heavy human steering; blockers shift to code review and verification, and the project ships internally while small inconsistencies remain.
AI-written summary · about
chapters (8)
- 0:00Introduction by Gene Kim
- 1:18LaunchDarkly Background
- 3:05The Haunted Codebase Problem
- 5:21Project Constraints and Goals
- 7:24Making the Codebase Agent-Ready
- 9:57Rewrite Reality vs. the Plan
- 13:14The Autonomy Trap
- 16:02Lessons Learned and What's Next
related talks
16mCoding Reinvented How AI Agents Are Shaping the Future of Software Development
24mBeyond Code Completion: The Shift of AI Programming from Pair to Peer
26mAI Agents as Your Co-Developer
13mDeveloping with AI: Balancing Speed and Code Quality
26mOrchestrating AI Agents for Large-Scale Refactoring
30mBuilding Tools with Tools: Modern AI-Driven Development
25mPanel The State of Web Development in 2026
14mAgent-Driven Software Development