AutoResearch does what the name suggests: give it a research question, it runs an autonomous process — searching, reading, synthesizing — and produces structured output. Minimal dependencies. Readable code. Actually useful.
Why "Minimal" Matters
The AI research agent space is cluttered with complex frameworks. Karpathy's explicit goal: a repo small enough to read in an afternoon, clean enough to modify without breaking, good enough to use for real work.
This reflects his recurring philosophy: code you can understand is code you can trust, debug, and extend.
What It Does
AutoResearch runs a loop:
- Generates search queries from your question
- Fetches and reads sources
- Synthesizes findings
- Identifies gaps, generates follow-up queries
- Repeats until confident
- Produces structured output
The agent does what a good human researcher does, in the same order, with the same verification steps.
The Next Step: Async Collaborative Agents
Karpathy has described the direction: making AutoResearch asynchronously massively collaborative — multiple agents working different threads of a question, sharing findings, building on each other's work. An agent research team, not a single agent.
The minimal repo is the starting point. For solo founders doing recurring competitive research or technical deep-dives, it's worth running now.