a16z · Oct 16, 2026 · 1h 0m · 53m saved
With Jad Naous, CEO of Grepr
A recent feature that used to take a single engineer about six months took Jad Naous ten days with AI — a factor of ten to twenty speedup on real production code.
Yet the same engineer who built that feature says he still won't let AI run a live, distributed production cluster unsupervised.
"That took maybe about 10 days. In previous days, that would have taken maybe about six months for a single person. It is insane... sort of a factor of 10, factor of 20 improvement."
Jad built architecture at AppDynamics, scaling it from roughly 200 to 2,000 employees before its Cisco acquisition, then ran product at Imply before founding the observability startup Grepr in 2023.
The full interview is covered here so you can skip it. 1 hour of audio, 7 minutes of reading.
Here are the 4 lessons that matter.
This Scuttlebutt Summary is available in full to premium subscribers. If you're not one already, please consider subscribing to keep reading.
Key Takeaways
A recent solo feature took 10 days with AI versus an estimated six months before it
A factor of 10 to 20 speedup on real production code
AI amplifies existing skill rather than leveling the field — stronger engineers pull further ahead, weaker ones see less gain
AI is strong at syntax, weak at semantics — what a system actually does at runtime, which isn't visible from the code alone
Operations is a heavier-tailed problem than writing code, and Jad doesn't think AI is ready to run a cluster unsupervised
1. Ten Days, Not Six Months
Jad runs a roughly 10-person engineering team at Grepr and described how AI coding tools evolved from simple autocomplete into something good engineers learn to "train" through their own workflow choices — which harness, which model, how much sub-agent parallelism to allow. He now runs Claude Code, via WSL on Windows, with Opus 5.5, often working three to four parallel agent sessions that communicate directly through the command line rather than through a shared pull request, with one agent acting as a "tech lead" reviewer for another's implementation before Jad personally reviews the combined result.
A recent project to add "auto-promotion" — automatically shredding semi-structured log fields into queryable Parquet/Iceberg columns in Grepr's data lake — took about 10 days solo
"That took maybe about 10 days. In previous days, that would have taken maybe about six months for a single person. It is insane... sort of a factor of 10, factor of 20 improvement." — Jad Naous
He also described the mechanics of running several agents at once.
"You basically tell Claude to spin up command lines of agents because those maintain their persistent session over time... you can actually run them from other sessions." — Jad Naous
