---
title: Autoresearch for all
pageLayout: default
slug: posts/Autoresearch-for-all
permalink: https://stateless.computer/posts/Autoresearch-for-all.md
generator:
  quartz: v4.5.2
  hostedProvider: Cloudflare
  baseUrl: stateless.computer
full: https://stateless.computer/llms-full.txt
---
The core of autoresearch lies in the idea that you can design a loop that will iterate in and on itself to improve a well-defined objective.

Seems simple enough at first but you will encounter two token-hungry challenges:
a) How do you define your goal in the way the system understands
b) How do you teach the system to _improve_

I got to explore autoresearch as an interview assignment for a company. I had heard of autoresearch before but this was my first implementation-level dive into it.

Needless to say I designed a very average loop as my first but I had some ideas which couldn’t make it to the assignment so they could end up on this blog here. This piece is written by me and so is the code. Any errors or shit code you spot is _proudly all mine_.

## Mental model for autoresearch

Just like any other agent like a code review agent or a pentest agent system, the autoresearch loop must also be designed by taking primitives from the human way of thinking and research.

### How to do research?

## The loop in the room

// elephant in the room address mee meme

Fundamentals should be correct. The tooling usage and observability is the main thing. The quality of the loop is directly dependent on the quality of _done_.

## Multi-agent design

## Enforced steps

## Novelty judge

## Local minimum