AUTORESEARCH, EXPLAINED · 2026-07-15

What is autoresearch?

Autoresearch is the use of autonomous AI agents to run the machine learning research loop itself: proposing experiments, writing the code, training and evaluating models, and deciding what to try next, while humans set the goals.

Why it exists

The motivation is simple: training a neural network is hard. There are many levers and knobs to turn, many hacks and tricks you need to know, and the design space is huge. You have to run many experiments before you find the right combination, and often you do not even know what the design space is.

Autoresearch automates that search. AI agents are already good at writing training code, and the LLMs behind them have been trained on a lot of data, including many of the papers and articles describing the hacks that make neural networks work. Given structured feedback about what happened during a run (losses, gradients, evals, logs) they can also decide what to try next. That turns one researcher's goal into a self-correcting loop of experiments.

Where the idea comes from

At its core, autoresearch is simple: a prompt that tells an AI agent to iterate on a task, plus an eval that scores every attempt. The name went mainstream in March 2026, when Andrej Karpathy released a repo with exactly that setup[1]: a single GPU LLM training script of about 630 lines (a stripped down nanochat core) plus a Markdown prompt. The agent reads the prompt, edits the training script, trains for 5 minutes, keeps the change if validation loss improves, and repeats overnight.

But the idea of using LLMs to automatically train models and discover algorithms has a longer history and a research community behind it. People have been thinking about this for a while: among other early works, DeepMind's FunSearch put LLM-driven program search in Nature back in 2023[3]. AlphaEvolve scaled the recipe in 2025 and found, among other things, a faster way to multiply matrices, the first improvement over Strassen's algorithm in 56 years[2]. Sakana AI's ShinkaEvolve open sourced a sample-efficient take on the same loop a few months later[4]. Some of the results:

Andrej Karpathy's X post introducing the autoresearch repo, with a chart of 83 experiments and 15 kept improvements
Karpathy announcing autoresearch, Mar 7, 2026. Click to open on X.

How agentic model training works

Agentic training of models means the agents drive the full training loop, not just a search over hyperparameters. In a typical autoresearch run:

Autoresearch vs. AutoML

APPROACHWHAT IT DOESWHAT IT CAN'T DO
AutoML / HPOSearches a predefined space (hyperparameters, sometimes architectures) with fixed strategies.Can't write new code, change the data pipeline, or pursue an open-ended research goal.
AutoresearchAgents write real code in your repo, run and score experiments on your infrastructure, and iterate toward a goal.Doesn't replace judgment: humans still choose the goals and the evals that define success.

What an autoresearch platform provides

Running hundreds of agents against one goal takes more than a prompt. An autoresearch platform like Autolab supplies the infrastructure layer:

What teams use it for

Autoresearch applies wherever a number needs to move and experiments are the way to move it: pre-training and fine-tuning of language models, post-training recipes (SFT, RL, distillation), computer vision and robotics model training, and inference optimization: quantization, kernels, and serving configs tuned for latency and cost.

Start running autoresearch today

Autolab installs in one line, points at the eval script that prints your metric, and starts the agents on your own compute.

$ curl -fsSL app.autolab.ai/install.sh | sh

References

  1. [1] Karpathy, A. autoresearch: AI agents running research on single-GPU nanochat training. GitHub repository, March 2026. github.com/karpathy/autoresearch; announcement: x.com/karpathy
  2. [2] Novikov, A. et al. AlphaEvolve: A coding agent for scientific and algorithmic discovery. Google DeepMind, 2025. deepmind.google
  3. [3] Romera-Paredes, B. et al. Mathematical discoveries from program search with large language models. Nature 625, 468-475, 2024. nature.com
  4. [4] Lange, R. T. et al. ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program Evolution. Sakana AI, arXiv:2509.19349, 2025. arxiv.org/abs/2509.19349