The End of Coding: Andrej Karpathy on Agents, AutoResearch, and the Loopy Era of AI (AI column)

I saw it. podcast "No Priors" with Andrej Karpathy. This series was published. 20 March 2026 Sarah Guo, founder of Conviction and host of the podcast, talks to Andrey, and the main idea of the release was that software engineering was already shifting from writing code to orchestrating agents. According to Andrey, since December, he almost stopped writing code with his hands and sharply shifted to the mode of delegation of tasks to agents; in his opinion, many engineers at the frontier level of default workflow had already radically changed.

Another interesting thought is that the bottleneck has shifted from the IDE and even computing power towards the human side - Andrei describes this as a transition from thinking about FLOPS to thinking about the bandwidth of tokens. (token throughput). If the agent can autonomously 20 When you do a piece of work, the value of the engineer shifts into the correct decomposition of tasks, running multiple threads in parallel, and reviewing at the architecture level rather than lines of code.

If a task has objective metrics, the agent can change the code, run short experiments, measure the result and leave only improvements. In the open repositories Karpathy is designed almost as a minimal research OS: the agent rules train.pyIt works in a fixed manner. 5Minute budget and optimize val_bpb; program.md In fact, it becomes the code of the research organization. This is his automatic work in the cycle. experiment → metric → change → repetitionWhere no man is needed.

However, Andrei notes that such a regime does not always work, but only where there is a verifiable reward and a clear way to assess the result. (eval). Where taste is needed, there are informal expectations and soft quality criteria, the models are still very inconsistent: in one place shiny, in another - surprisingly naive. This is an important amendment to the hype about “autonomous developers”.

Andrey also spoke about his home agent Dobby, who manages lights, HVAC, curtains, swimming pool, spa and home security through WhatsApp. From this he draws a stark conclusion: perhaps a noticeable part of today’s applications is a temporary UX layer, and the more correct architecture of the future is an API/tool surface + agent as glue.

All in all, it's a really interesting podcast in which Andrei shares practice rather than futuristic ideas.

#Engineering #AI #Metrics #Software #DevEx #Productivity #DevOps #Architecture #Culture #Engineering #SystemDesign