CharlieKey takeaways from your favorite podcasts
← 16 October 2026a16z

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

Jad argues AI already transforms production coding and can plausibly handle operations by running debugging workflows on deterministic change signals, while Martin doubts ops fits the same distribution and warns about real-world limits.

Watch on YouTube ↗

Charlie's summary

  1. Observability spend is a proxy for reliability, and its marginal reliability per dollar has flattened

    Jad found engineering leaders spending 5 to 20% of infrastructure cost on observability, but what they wanted was reliability. With microservices, data volume grew faster than output, so each extra dollar bought less. His answer: extract signal upstream in the data stream, cutting cost by 90% and feeding those signals to AI that troubleshoots and fixes problems.

    ▶ 7:06
  2. AI widens the gap between good and average engineers rather than closing it

    Jad says the best engineers put AI into their workflows so that it trains them in how to talk to it. As the team grew, the disparity between okay and really good engineers multiplied 10x. Good engineers just became better ones; weak ones did not suddenly become good. Jad ties this to his distributed-systems work, where the engineer's brain must handle complexity the models cannot. His team is small, so he hasn't tested this across 100 people.

    ▶ 13:13
  3. Jad's multi-agent workflow compressed six months of production work into ten days

    Jad runs Claude Code with Opus 5.5 and tells it to spin up command-line sub-agents, which keep persistent sessions. A tech-lead agent reviews the PR and design doc, and Jad checks every algorithm and PR boundary before telling it to start. He runs three or four sessions at once. Reviewing shifted toward tests and golden expectations. The auto-promotion (shredding) feature for a data lake took about 10 days versus roughly 6 months alone, a 10 to 20x gain.

    ▶ 20:19
  4. AI can optimize against benchmarks but cannot know production facts like memory sizing without experiments

    When Jad's model generated unoptimized SQL-derived code, it fixed it after he pointed out repeated terms, because it could ground hypotheses in performance benchmarks and real query data. But asked how much memory to give a system, it has no idea; that must come from production experiments. Jad says the models are getting good at recognizing what they can't answer and running an experiment to find out.

    ▶ 28:01

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.