At a glance
WHAT IT’S REALLY ABOUT
AI writes code fast, but ops needs change-detection workflows
- Jad Naous traces Grepr’s founding to a widespread push to cut infrastructure spend, where engineering leaders repeatedly singled out observability costs as the most painful line item.
- He argues observability spend ballooned as microservice architectures increased telemetry volume, while reliability improvements per additional observability dollar plateaued.
- Naous describes how AI has moved from autocomplete to a structured, multi-agent development workflow that can produce production-grade infrastructure features dramatically faster—especially for engineers who can precisely specify constraints and validate results.
- On operations, he contends AI struggles to “predict” runtime behavior directly but can succeed when paired with deterministic change detection, then guided through human-like debugging workflows and experiments.
- He contrasts operations with security, arguing security is harder because an adversary intentionally searches for rare states, making purely data-driven coverage more fragile.
IDEAS WORTH REMEMBERING
5 ideasAI’s big gap is moving from writing code to managing running systems.
Naous argues that code generation is easier because the “syntax manifold” is well-covered by training data, while operations requires reasoning over runtime behavior, changing workloads, hidden human context, and heavy-tailed failure modes.
Observability became expensive because telemetry exploded and marginal reliability returns collapsed.
Grepr’s origin story is cost-driven: microservices massively increased telemetry volume, making observability spend rise faster than delivered functionality, while reliability gains per extra dollar flattened.
AI amplifies engineer skill gradients rather than equalizing them.
Naous reports that strong engineers get disproportionately stronger with AI because they can decompose problems, specify constraints, and validate outcomes; AI doesn’t remove the need for systems-level understanding.
High-leverage AI coding is a managed, multi-agent pipeline—PRD first, code second.
He describes a workflow using Claude Code (Opus 5.5) with multiple persistent command-line “sub-agents,” starting with a PRD/design doc, then parallel implementation, followed by human re-integration and review.
For well-specified infrastructure work, AI can compress months of effort into days.
A concrete example: implementing “auto-promotion/shredding” in a data lake (schema evolution, metadata commits, writer coordination) took ~10 days solo, which he estimates would have taken ~6 months previously.
WORDS WORTH SAVING
5 quotesThe dirty secret for the longest time of running, like, a distributed cluster was just human beings managing it.
— Martin Casado
There, there, there's, the last PR that I worked on, or, like, the last set of PRs, was what we call auto, auto promotion in a data lake... that took maybe about 10 days. In previous days, that would've taken maybe about six months... for a single person.
— Jad Naous
The goal isn't to like have the model just be like, "This is what happened." It's, "Hey, run through this workflow to determine what happened."
— Jad Naous
The world that the AI lives in is all digital, and there's the actual world. And, like, those things are not necessarily the same.
— Martin Casado
Self-driving cars is around the corner... And it took 20 years and $170 billion of investment to get to worse unit economics than Uber, right?
— Martin Casado
High quality AI-generated summary created from speaker-labeled transcript.
