AGENTS.md

advancedwidespread

AGENTS.md is a simple text file in a software project folder that describes how coding AI agents should work on that project.

In plain terms

Imagine a new colleague arriving at your workplace with a sheet on their desk: where to find things, what the house rules are, what not to touch, and who to contact when finished. AGENTS.md is that same sheet, but intended for the coding agent, not a human.

The agent does not know on its own how testing is usually done in a project, in what style code is written, or which files must not be touched. If you write this once into this file, you do not need to explain it again every time.

What sets it apart from earlier solutions is that it is a common, open format. Previously, many tools expected their own named rule files, whereas AGENTS.md is read in the same way by agents from many different vendors.

How it works

The project owner creates a file named AGENTS.md in the project's main folder and writes the information in everyday language using simple Markdown formatting.
Typical content includes: how to start and test the program, which coding conventions apply, what to watch out for, and how to submit the finished work.
When the agent starts working, it finds and reads this file, and its content becomes part of the text it works from.
In large projects, subfolders can have their own AGENTS.md files. If two files' rules conflict, the one closer to the file the agent is currently working on counts. The user's direct instruction overrides both.
The file changes and can be fixed just like code: if the agent regularly makes a mistake in something, the team writes in a new rule.

An everyday example

A web shop development team noticed that the agent regularly forgot to run tests before marking work as complete. They added to AGENTS.md: "run tests after every change and submit only if the result is error-free." From then on, the agent mostly worked this way on its own, while the team verified that tests actually ran.

Who is affected

developerIt writes and maintains the file, and this determines how usable the work received by the coding agent will be. → Code

entrepreneur, executiveCompany development rules can thus be recorded in one place, verifiably, without binding the team to a single vendor's tool.

parent, student, teacherThose learning or teaching programming with AI assistance can specify in this file that the agent should explain rather than just provide ready-made solutions. → Education & Family

A common misconception

MythWhat you write in AGENTS.md is guaranteed to be followed by the agent, as if it were program code.

In realityThe file is a guide, not a lock: the agent takes it into account, but may ignore individual rules if the text is long or contradictory. Real prohibitions must be solved with permissions and controls.

Limits and risks

Never include passwords, access keys, or other secrets, because the file spreads with the code and anyone with access to the project can read it.
An AGENTS.md file from an external project may contain malicious instructions, so check it before allowing an agent to run on code downloaded from an unknown source.
Long, outdated, or self-contradictory files degrade results, so they require regular maintenance.
Not every tool reads files automatically: some agents expect files with their own names, requiring a reference or separate configuration.

Similar, but not the same

README – The README introduces the project to humans, while AGENTS.md contains practical work instructions for the agent.
Rendszerprompt – The system prompt is written by the tool creator and applies to every project. The AGENTS.md is written by you and applies only to that specific project.
CLAUDE.md – The CLAUDE.md file serves the same purpose but is used only by a single tool, Claude Code. In contrast, AGENTS.md is manufacturer-independent. A CLAUDE.md file can reference an AGENTS.md file, allowing both to be used together.

Related concepts

Broader

AI-agent, Context engineering, Open standard

Narrower

–

On the edge

coding agent, Rendszerprompt, Model Context Protocol, Agent skills, Prompt injection

A brief history

The AGENTS.md format emerged in the summer of 2025 as a joint, open standard. Several companies developing coding agents worked on it together, including OpenAI, Google, Cursor, Factory, and Sourcegraph. In December 2025, it came under the care of the Agentic AI Foundation established within the Linux Foundation.

Where things stand

At the end of 2026, most coding agents read it, and tens of thousands of open-source projects use it. No official standards body has formalized it, but in practice it is the common solution, although some tools still use their own file name.

Status: 2026 November

In the news

Appears in 5 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Malicious files in the Downloads folder can hijack AI coding agents, researchers sayOct 26, 2026 · Code
Planted AGENTS.md file in Downloads folder hijacks AI coding agentsOct 25, 2026 · Code
Claude Code's AGENTS.md support reportedly forced users to choose between privacy and the featureSep 24, 2026 · Code
Claude Code now reads AGENTS.md, and auto mode safety checks are freeSep 21, 2026 · Code
Anthropic launches Claude Code Projects beta and merges Claude chat with CoworkSep 16, 2026 · Agent Age

Also known as

AGENTS.md · agents.md · Agents.md file · AGENTS.md standard · AGENTS.md format · Agent instruction file · agent instructions file · agent file

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.