AGENTS.md and code agents: the developer's new tooling

From autocomplete to agent

In just a few years, coding assistance has changed in nature: from line-by-line suggestions, we've moved to agents capable of reading a whole codebase, running tests, opening a pull request. The problem is no longer "generating code," but giving the agent the right context — build commands, conventions, architecture.

AGENTS.md: an open standard, in Markdown

That's exactly what AGENTS.md solves: a simple Markdown file, at the root of the project, that explains to any agent how to work here — build and test commands, code style, frameworks, architectural decisions. No imposed schema, no tool to install: any agent that reads it benefits immediately.

The format was formalized in August 2025, then donated to the Linux Foundation's Agentic AI Foundation in December 2025. By the end of 2025, more than 60,000 open-source projects had adopted it, and more than twenty tools support it — GitHub Copilot (native support since August 2025), OpenAI Codex, Cursor, Gemini and Jules, Windsurf, Zed, and others. A sign of the times: Next.js 16.2 now generates an AGENTS.md in create-next-app.

Why it's more than a fad

An agent with no context rediscovers the project every time: it guesses the commands, ignores your conventions, produces code you have to rework. A well-written AGENTS.md removes that discovery phase: the generated code respects your standards from the start, and review round-trips shrink.

It's basically the same idea as good onboarding documentation — except here the newcomer is a machine that reads fast and forgets nothing.

My position

I say it often: AI is part of my tools, not my foundations. A well-equipped agent speeds up research, prototyping, mechanical tasks. But it's the developer who writes the AGENTS.md, who decides the architecture, who reviews and validates. The tool doesn't replace judgment: it amplifies it.

An agent is only as good as the context you give it. AGENTS.md is that context, versioned with the code.