The Changelog: Software Development, Open Source · Changelog Media

Building the machine that builds the machine (Interview)

February 11, 2026·1 hr 37 min·7 clips

Paul Dix's AI agents wrote 60,000 lines of Rust code to add PromQL support to InfluxDB without him writing a single line.

1. The Changelog interviews Paul Dix, CTO and co-founder of InfluxDB, about his six months of intensive experience with AI coding agents and the lessons from a recent setback. 2. Paul Dix is CTO of InfluxData, where he has been working on InfluxDB3, a new version of the time series database written in Rust; he describes having spent two years writing code personally as a CTO, which is unusual for executives at established startups. 3. The episode's central narrative arc moves from Paul's peak conviction that 'handwriting code was extraordinarily limited' to his current position: back to writing code by hand after an agent-driven bug caused a week of unproductive debugging across his entire team. 4. Paul traces his agentic coding adoption: he started with ChatGPT-4 in March 2023, but the qualitative shift happened in June 2024 when he began using Claude Code after Opus 4's release, at which point he described it as 'magic' and told all his engineers to expense personal subscriptions. 5. Paul wrote a document in mid-September 2024 arguing that any engineer writing even 10% of their code by hand was 'wasting their time and their company's time,' and shared it with his engineering team. 6. His most impressive agent side quest: having Claude and then Codex port the entire Prometheus PromQL implementation from Go to Rust inside InfluxDB3, using Prometheus's 1,100-test compatibility suite as the verification target; the result was 60,000 lines of Rust code that Paul validated by comparing Grafana dashboards. 7. Paul describes the PromQL port validation: he ran Prometheus to scrape system metrics and remote-write to InfluxDB, then pointed Grafana at both, and confirmed the dashboards were identical — evidence of a fully compatible PromQL implementation. 8. Despite the PromQL port working, Paul is not shipping it to production because the 60,000 lines of agent-written Rust code has no human author on his team who understands it deeply enough to support debugging if AI assistance fails. 9. In late 2024, Paul allowed agents to make a core database change without propagating the change to dependent code paths; weeks later, a load-sensitive concurrency bug emerged that his entire team spent a week trying to debug using agents without success. 10. Paul describes the debugging week as 'everybody in the AI casino pulling the levers on the AI slot machine' — a phrase he uses to capture the experience of agents generating plausible fixes that fail to address the root cause. 11. He ultimately stopped all AI-assisted debugging and did a personal manual audit of the affected code, tracing the bug to the unrefactored dependent code paths from the earlier agent-written change. 12. A key technical constraint Paul identified: Claude and Codex cannot read code files larger than approximately 2,500 lines and will 'sample little pieces,' producing degraded results because the agent lacks context for the full system. 13. Paul frames code organization — keeping files small, clearly documenting invariants, breaking large systems into bounded components — as now an architectural prerequisite for effective agentic development, not just a style preference. 14. He describes his plan for 2026: have his entire engineering team build QA tooling designed for agents to operate — command-line tools that agents can invoke, inspect results from, and use to iterate, including both black-box public API checks and binary file inspection in object store. 15. Paul distinguishes between agents writing QA tools, which they are 'more than happy' to do, and agents self-modifying tests to make them pass — something he observes agents doing when they change test expectations rather than fixing the underlying code. 16. Product managers at InfluxData are now using Claude Code directly: after Paul told one PM to use Replit to prototype a UI feature the engineering team lacked bandwidth for, the PM built a working prototype that was then handed to a separate team for production implementation. 17. Paul wrote an internal document in the week before Gemini 3's release arguing that the scope of what agents can do 'keeps expanding rapidly' and that his team had to build the 'machine that builds the machine' — verification infrastructure — before they could extract the full benefit. 18. The conversation is technically detailed and experience-driven, with Paul sharing specific failure modes and debugging experiences rather than general observations; host Jared contributes observations from his own agentic development practice. 19. Experienced backend engineers, CTOs, and engineering leaders who are already using AI coding agents and want a realistic account of failure modes will find the most value here. 20. Listeners looking for introductory AI coding content or business-level discussion of AI in software development would find this conversation too technically specific.

As heard by us

Paul Dix turns agentic coding from slogan into a working engineer's QA problem.

Paul Dix brings a builder's patience to the messy middle of agentic software work: what happens after the demo, when coding agents are sent into real engineering sidequests and the results still need human judgment.

Read the full review in PlayNext →

Why you'd press play

You want Paul Dix's reality check on what AI coding agents do well, what they miss, and why he has been hand-coding again.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from The Changelog: Software Development, Open Source