Karpathy Releases autoresearch: AI Agents Start Doing AI Research

In March 2026, Andrej Karpathy put a small open-source project called autoresearch on GitHub. By October, the Chinese outlet Jiqizhixin had made it the starting point for a long analysis of recursive self-improvement in AI: the situation in which models take part in building the next models. The repository is modest, but its framing is telling.

A Memory from the Future

The README opens not with installation instructions but with a short, stylized sketch "from the future." In it, frontier AI research is remembered as the work of "meat computers." They needed food and sleep, and from time to time they synchronized via "sound waves" in a ceremony called a "group meeting."

Karpathy is joking, but the joke is accurate. He describes the routine of a lab: hypothesis, code change, experiment run, reading the metrics, next hypothesis. Language models and coding agents can already perform every step of this cycle. Chain the steps into a closed loop, and the human gradually drops out of it.

Recursive Self-Improvement as an Engineering Problem

For a long time, recursive self-improvement was discussed as a philosophical scenario: one day a system becomes smart enough to improve itself, and from there everything goes exponential. The title of Jiqizhixin's piece, "When AI Starts Building AI," captures a different view. The process is broken down into narrow, verifiable operations, and each one is automated separately: writing code, running experiments, analyzing results.

Metrics, not intuition, set the scale here. An agent needs a clear goal, a fast verification loop, and protection against self-deception, where an improvement on paper does not translate into an improvement in practice. These conditions determine where research automation will work first and where it will hit a ceiling.

The available excerpt of the article reveals little:

• project author: Andrej Karpathy;
• name: autoresearch;
• platform: GitHub, open repository;
• publication date: March 2026;
• framing: the human researcher as a historical stage, not a constant;
• source of analysis: Jiqizhixin, article dated October 10, 2026.

The excerpt contains no numerical results, benchmarks, or description of the architecture. It gives no basis for judging the project's effectiveness; only the stated framing can be assessed.

Why This Matters Right Now

What makes Karpathy's project interesting is that it is small. Research automation is usually associated with closed labs and huge clusters. An open repository that you can read in one evening makes the topic practical: any developer can run such a loop on their own task and see where it breaks.

The questions remain the same. Who checks the agent's decisions if the checker is also a model? How do you avoid optimizing the metric instead of the goal? How much compute goes into experiments that produce nothing? The excerpt offers no answers, and that is fine: the project sets a frame rather than closing the topic.

What Teams Should Do

If you build ML products or internal research pipelines, it is worth breaking your experiment cycle into stages and marking which of them can be formalized: a fixed run budget, an unambiguous metric, reproducibility. These are the stages to hand to agents first. The rest need a human reviewer.

It is also worth reading the repository itself and the full Jiqizhixin article. It likely contains implementation details and examples that did not make it into this excerpt. Conclusions about how strong the approach really is should only be drawn after testing it on your own tasks.