What is scientific-agents?

This repository holds a collection of 503 expert profiles, one for each scientific or engineering profession, designed to change how an AI coding agent reasons about problems in that field. Each profile is a plain Markdown file called AGENTS.md that describes how a senior practitioner in that discipline actually thinks: what databases and instruments they reach for, how they frame a question, what controls they apply, how they evaluate a claim, what mistakes they watch for, and how they report findings. Drop one into a project and the AI agent stops responding like a general-purpose assistant and starts reasoning in a way that someone in that profession would recognize. The profiles work across several popular AI coding tools without modification: Cursor, Claude Code, OpenAI Codex, Google Antigravity, and OpenCode all read AGENTS.md files from the project root. The format follows an open standard, so the same file can be reused across tools or shared across a team. For Claude Code specifically, you create a CLAUDE.md file that imports the AGENTS.md profile, since Claude Code's native filename is different. The profiles are built through a research process rather than written as generic instruction prompts. Before each one is drafted, the team gathers field-specific information across multiple dimensions: foundational theory, how experts in that field define a problem, relevant tools and data sources, statistical methods, known failure modes, communication conventions, and any relevant regulations or ethics requirements. The goal is to collect specific, verifiable details rather than broad statements that could apply to any science. A finished profile is tested by checking whether swapping in a different profession's name would make the sentences still read as true. If they do, those parts are cut or made more specific. Examples of included professions mentioned in the README include developmental biologist, clinical epidemiologist, bioinformatician, and radio astronomer. The full catalog is indexed in a catalog.json file in the repository. The full README is longer than what was shown.