Karpathy’s AutoResearch case: remove the human bottleneck, keep the verification

On 20 March 2026, No Priors published a long-form conversation in which Andrej Karpathy described a sharp change in his own engineering practice: less hand-written code, more direction of several coding agents, and a growing effort to keep useful work running without waiting for the next human prompt.

Andrej Karpathy Watch analysis

What happened

Karpathy said his personal balance shifted after December from mostly writing code himself to mostly delegating larger pieces of work. His picture of mastery is not one clever chat window. It is a set of agents handling separate jobs, with the person choosing tasks, reviewing results and improving the instructions around them.

AutoResearch is his clearest test of that idea. He framed a small language-model training problem with an objective, a metric and explicit boundaries, then let an agent repeatedly propose changes, run experiments and keep improvements. He reported that an overnight run found useful hyperparameter adjustments he had missed, despite years of experience tuning similar systems.

Why it matters

The important shift is from generating an answer to operating a verifiable loop. That can multiply experimentation where success is cheap and objective to test, such as runtime or validation loss. Karpathy also supplied the brake pedal: if a result cannot be evaluated reliably, it is a poor fit for this kind of autonomous optimisation.

His account was candid about the rough edges. Agents can be startlingly capable and then plainly wrong, and he had not granted broad email or calendar access because of security and privacy concerns. This is a practitioner’s field report, not proof that supervision has become optional.

Our read

Natural language looks increasingly like a control surface for software, not a replacement for the machinery beneath it. The engineering moves into objectives, permissions, tests and observability. Remove the human from every loop without building those first and the shed has merely invented a faster way to misplace the spanner.

What to watch

• Independent reproductions of the reported AutoResearch gains.
• Evaluation methods for work without clean numeric targets.
• Permission and audit controls for persistent agents.
• Whether multi-agent supervision becomes ordinary engineering practice.

Discussion spark: Is natural language becoming the main programming interface, or is it a new control layer over conventional software and tests?

Independent WittyWires tracker for public updates about Andrej Karpathy. Not affiliated with or endorsed by Andrej Karpathy; this is not an official account.