Random · Agents
MathModelAgent in nine questions
The agent behind the Bitcoin paper, explained the way people actually asked about it, including what its 2026 release changed.
What is it?
A recipe for turning a modelling question into a finished paper: analyse, model, code, draw, write, verify. Built by jihe520, a computer-science student in Wuhan, for the competitions where a team gets a problem on Friday and hands in a paper on Monday. About 5,000 stars on GitHub.
Is it a model?
No. It has no weights. It is six plain-text skills: what to ask, what to produce at each stage, what is forbidden. The intelligence is whichever coding agent reads them. The official desktop app bundles Claude Code; I run the same skills on Hermes.
What changed in 2026?
Everything under the hood. The 2025 version was an app: a web server, four agents handing work to each other, Redis, a sandbox. In 2026 the author dropped all of it (his words: no more harness layer) and kept only the method as text any capable agent can follow, plus a desktop wrapper so a student installs nothing else.
What does it produce?
A plan, an analysis report with the model spelled out, runnable code with a results file per run, figures, a roadmap diagram, a Typst or LaTeX paper from 17 competition templates, and a verification report.
How is it different from asking a chatbot?
The order and the gate. It must resolve ambiguities before modelling, run code before writing, and it may not put a number in the paper that is not in a results file. The last stage checks that in nine steps and fails the run.
How do I install it?
One line: npx skills add jihe520/MathModelAgent. Then /1start-mathmodel and answer four questions. Or download the desktop app (macOS signed, Windows not yet) and add one model API key.
What did it do for me?
The Bitcoin cycle paper on this site: five model families, a backtest on every top and bottom since 2013, sixteen pages, 235 automated number checks on the final text, in one afternoon.
Where does it fail?
It is prescriptive, not clever: it fits a line through three points because the skill says to. The gate checks consistency, not meaning; on the Bitcoin paper it passed an interval built from the wrong backtest rows and a human caught it. The prompts are in Chinese. Free for personal use only.
Should you use it?
If your question is 'model this, solve it, write it up', yes, and read every derivation it hands you. If your question needs taste before method, it will not supply the taste.