# AutoResearch

> **TL;DR.** AutoResearch is an agent loop in which an LLM autonomously edits a single training file, runs a fixed 5-minute experiment, checks whether a chosen metric improved, and keeps or reverts the change — repeating overnight.

- **Category:** AI / Developer Tools / Agent Research
- **Stage:** rising
- **Age:** 101 days
- **Origin date:** 2026-03-07
- **First detected:** 2026-04-20
- **Canonical URL:** https://earlyterms.com/term/autoresearch
- **Sources:** 8 primary URLs

## Definition

AutoResearch is an agent loop in which an LLM autonomously edits a single training file, runs a fixed 5-minute experiment, checks whether a chosen metric improved, and keeps or reverts the change — repeating overnight. It is a minimal "modify → verify → keep/discard → repeat" harness, not a heavyweight framework.

The name crystallized when Andrej Karpathy open-sourced [karpathy/autoresearch](https://github.com/karpathy/autoresearch) on [March 7, 2026](https://github.com/karpathy/autoresearch), pointing it at his own nanochat GPT-2 codebase. Two days, ~700 experiments, ~20 real improvements, and an 11% end-to-end speedup later, the repo had 74k+ stars and the term had been adopted as a generic category for autonomous experiment loops.

## Example

Karpathy's reference setup uses three files with strict ownership: `prepare.py` is immutable and handles data plus the `val_bpb` evaluator, `train.py` is the agent's sandbox, and `program.md` is the human-written research brief. The agent proposes a change (e.g. a QK-norm scaler, a banded-attention tweak, an AdamW beta), trains for exactly 5 minutes, and git-reverts anything that doesn't lower validation loss.

## Analogy

Like leaving a junior researcher alone overnight with a stopwatch and one dial — they try things, and only the wins survive.

## Why it's emerging now

Karpathy open-sourced AutoResearch on March 7, 2026 — 630 lines of Python that let an AI agent run ~100 training experiments a night on a single GPU. Two days of autorun shaved 11% off his already-tuned nanochat GPT-2 pipeline, and Fortune ran the "loopy era" thesis ten days later.

## Related terms

- *related:* nanochat
- *parent:* agent loop
- *related:* agent harness
- *competitor:* AutoML
- *child:* AutoResearchClaw
- *child:* AutoKernel
- *related:* parallel agents
- *related:* coding agents
- *related:* Claude Agent SDK
- *related:* managed agents
- *related:* agentic coding

## Sources

1. [karpathy/autoresearch — canonical repo](https://github.com/karpathy/autoresearch)
2. [VentureBeat — Karpathy's open-source autoresearch](https://venturebeat.com/technology/andrej-karpathys-new-open-source-autoresearch-lets-you-run-hundreds-of-ai)
3. [Fortune — Why everyone is talking about Karpathy's autonomous AI research agent](https://fortune.com/2026/03/17/andrej-karpathy-loop-autonomous-ai-agents-future/)
4. [DataCamp — Guide to AutoResearch](https://www.datacamp.com/tutorial/guide-to-autoresearch)
5. [Hacker News — launch thread (208 pts, Mar 7)](https://news.ycombinator.com/item?id=47291123)
6. [Hacker News — "Autoresearch on an old research idea" (428 pts, Mar 23)](https://news.ycombinator.com/item?id=47493460)
7. [SkyPilot blog — Scaling Karpathy's Autoresearch to a GPU cluster](https://blog.skypilot.co/scaling-autoresearch/)
8. [awesome-autoresearch — ecosystem list](https://github.com/yibie/awesome-autoresearch)

---
_Generated by EarlyTerms · https://earlyterms.com/term/autoresearch_