Paul Dix

AI Summary

→ WHAT IT COVERS Paul Dix, CTO and cofounder of InfluxDB, shares his six-month journey using AI coding agents like Claude and Codex to generate hundreds of thousands of lines of code. He discusses successful side quests including a 60,000-line PromQL implementation, the challenges of code review at scale, and why verification tooling matters more than code velocity in production environments. → KEY INSIGHTS - **Agent-Generated PromQL Implementation:** Dix used Claude Opus and GPT Codex to port Prometheus's PromQL query language from Go to Rust, producing 60,000 lines of code that passed 1,100 integration tests. The agent used Prometheus's existing test suite as verification signals and the Go implementation as a reference, demonstrating that agents perform better when given clear validation frameworks and reference implementations to follow. - **Verification Over Velocity:** Engineers can now produce 100 times more code than before, but code review becomes the bottleneck. Organizations need comprehensive QA suites that agents can execute autonomously, including black-box API tests, binary file inspection tools, and cloud-based validation environments. The verification suite must be more extensive than traditional unit tests, resembling what QA engineers historically built for comprehensive system validation. - **Code Organization for Agents:** Agents struggle with files exceeding 2,500 lines because they only sample portions rather than reading complete context. This leads to suboptimal decisions and architectural problems. Teams should break code into smaller, well-defined modules with clear invariants, making it easier for both humans and agents to maintain context. File size limits become critical constraints for agent effectiveness. - **Two-Person Team Structure:** The optimal team size for agentic development is two to three people, primarily two. One person can now manage 15 agents simultaneously, producing what previously required 100 engineers. The ideal pairing is a product manager and an engineer, where most work involves defining problems clearly, structuring solutions architecturally, and iterating on user experience rather than writing code. - **Production Deployment Challenges:** Despite generating hundreds of thousands of lines of functional code across multiple side quests, Dix ships none of it to production. The gap between demoable prototypes and production-ready software remains significant. Key barriers include performance optimization, support responsibility for AI-generated code, and lack of human familiarity with the codebase when agents cannot solve problems independently. - **Internal Tooling Adoption Strategy:** Dix mandated all engineers use Claude Code in June 2024, instructing them to expense personal subscriptions and upgrade without worrying about token limits. This removed psychological barriers to experimentation. By year-end, most engineers adopted agentic tools, but organizational processes remain unchanged with traditional pull requests, human code review gates, and sequential development workflows that limit velocity gains. - **Agent Ergonomics Over Developer Experience:** Database and infrastructure products should optimize for agent usability rather than developer ergonomics. This means comprehensive markdown documentation that agents can consume, command-line tools designed for programmatic access, and security features that lock down agent permissions. Organizations with good documentation, CLIs, and APIs positioned for agent consumption will see adoption advantages as agents become primary software creators. → NOTABLE MOMENT Dix describes letting agents work unsupervised for weeks, then discovering fundamental architectural problems requiring manual refactoring. Multiple engineers spent days using AI tools to debug timing issues under load, essentially gambling in what he calls the AI casino, pulling slot machine levers hoping for solutions. He finally solved it by auditing everything manually, revealing that agents excel at generation but struggle with complex debugging. 💼 SPONSORS [{"name": "Fly.io", "url": "https://fly.io"}, {"name": "Namespace", "url": "https://namespace.so"}, {"name": "TigerData", "url": "https://tigerdata.com"}] 🏷️ AI Coding Agents, Software Engineering, Code Verification, InfluxDB, Rust Development, Developer Productivity, Agentic Development