Real Alpha
Watch mode

Full episode · 59 min

Why AI Can Write Code But Can’t Keep It Running (Yet)

Jad Naous argues that AI can write production code, but running live systems requires new ways to detect change, test hypotheses and manage complexity.

a16z Podcast · Martin Casado with Jad Naous

The best minutes

  1. 17:52AI cut a complex data-lake project from six months to ten days7 min
  2. 53:20Real-world AI takes longer when the environment is unconstrained6 min

Also worth watching

  1. 27:10AI query optimization needs benchmarks, not guesswork3 min
  2. 10:48Experienced engineers get more leverage from coding agents4 min

Skip the Stanford and career history, the semantics discussion, and the extended debate on operational limits, security and venture capital.

Summary · 2 min read

Naous studied at Stanford, worked on OpenFlow and hardware security, and later did application-security research at MIT. He worked at AppDynamics and a16z, then became Ampli’s chief product officer before founding Grepr.

Naous says AI agents have increased the leverage of strong engineers, who can break down complex work, review designs and direct multiple agents. He says one data-lake project that might once have taken a single person six months took him about 10 days. For operations, Naous argues that AI needs useful signals from production, then must inspect changes, form hypotheses and run experiments; Casado stresses the difficulty of rare states and hidden operational knowledge. They agree that operating systems is harder than generating code and remains unresolved, while constrained digital tasks may be easier to automate than problems shaped by the real world.

Minute by minute

  1. 10:06

    AI widens the engineering gap. Naous says coding tools moved quickly from autocomplete to more capable agents. Engineers who learn to break down work and use different models and harnesses gain more leverage; as his team hired, he says the gap between okay and very good engineers multiplied 10x.

  2. 17:52

    Agents need human direction. Naous describes running several agent sessions at once, asking one agent to implement while another reviews, and having agents produce design documents before coding. He checks the algorithms, task boundaries and pull-request plan, then reviews the assembled work himself.

  3. 22:42

    A six-month project took ten days. Naous says he built an auto-promotion feature for a data lake, which turns fields in semi-structured log events into independently filterable columns. The work involved metadata, commits and coordination across writers; it took him about 10 days, compared with an estimate of six months for one person.

  4. 24:42

    Production behavior needs experiments. Casado points out that workload and access patterns shape a system’s correctness and performance in ways code alone may not reveal. Naous says benchmarks and representative query data can help an AI optimize queries, but it cannot know how much memory a system needs without experiments in production.

  5. 53:20

    Real-world domains take longer. Naous expects AI to advance fastest in constrained domains where solutions resemble the training data, and more slowly when solutions depend on the natural world. Casado cites self-driving as a long, expensive path and says changing systems to constrain the problem may help. Naous says customers he approaches often want Grepr’s telemetry-cost savings more than its AI operations features.

Published Oct 16, 2026.

More interviews