Earlier today, I saw Kath Korevec's (Google Lab's Director of Product) tweet about the fragmentation in AI agent file formats across tools like Claude Code, ChatGPT, and various frameworks while I was laying my son down for a nap. Everyone's using different approaches - some JSON configs, some plain text, some proprietary formats.
The problem: no standard way to define, share, or version AI agents across platforms.
So I built one, and two hours later it was live. The .agent format is essentially Markdown with YAML frontmatter - think Jekyll/Hugo but for AI agents:
I help debug Python code and suggest improvements...
```
Why this approach works:
Immediate compatibility - Any markdown parser handles .agent files
Existing tooling - Syntax highlighting, git diffs, pandoc conversion all work
Human readable - You can read/edit with any text editor
Structured metadata - YAML frontmatter enables tooling and discovery
Extensible - Add new metadata fields without breaking existing parsers
I've submitted the MIME type (application/vnd.agent+markdown) to IANA for official registration and built reference implementations in Python/JavaScript plus VS Code extension.
The repo includes working examples and a complete specification. The goal: make AI agent definitions as portable and interoperable as markdown files themselves.
Earlier today, I saw Kath Korevec's (Google Lab's Director of Product) tweet about the fragmentation in AI agent file formats across tools like Claude Code, ChatGPT, and various frameworks while I was laying my son down for a nap. Everyone's using different approaches - some JSON configs, some plain text, some proprietary formats. The problem: no standard way to define, share, or version AI agents across platforms. So I built one, and two hours later it was live. The .agent format is essentially Markdown with YAML frontmatter - think Jekyll/Hugo but for AI agents:
``` yaml --- agent: name: "Python Expert" capabilities: ["debugging", "code-review"] coding_languages: ["python"] ---
# Python Expert
I help debug Python code and suggest improvements... ```
Why this approach works:
Immediate compatibility - Any markdown parser handles .agent files Existing tooling - Syntax highlighting, git diffs, pandoc conversion all work Human readable - You can read/edit with any text editor Structured metadata - YAML frontmatter enables tooling and discovery Extensible - Add new metadata fields without breaking existing parsers
I've submitted the MIME type (application/vnd.agent+markdown) to IANA for official registration and built reference implementations in Python/JavaScript plus VS Code extension. The repo includes working examples and a complete specification. The goal: make AI agent definitions as portable and interoperable as markdown files themselves.