Open Agent Format

Version 1.2.0 | January 2026

An open, interoperable specification for describing agent composition

Embracing Open Standards

New Specification Notice

The Open Agent Format is a newly released specification. While we're working with major vendors to adopt this standard, most platforms haven't fully integrated OAF yet. We provide vendor-specific mapping guides below to help you convert between OAF and existing formats.

Core Design Principles

Filesystem as Truth

The directory structure and files define the agent - no hidden state or configuration

Interoperable

Works across Claude Code, Goose, Deep Agents, Letta, and other harnesses

Composable

Agents reference skills, packs, weblets, MCPs, and other agents

Version-aware

Built-in semantic versioning and version history support

Human-readable

Markdown for instructions, YAML for metadata - easy to read and edit

Harness-agnostic

Core format is independent of any specific platform

Key Features

Quick Example

A minimal OAF agent requires just one file:

my-agent/
└── AGENTS.md

The AGENTS.md file contains YAML frontmatter with metadata and Markdown instructions:

---
name: "My Assistant"
vendorKey: "acme"
agentKey: "assistant"
version: "1.0.0"
slug: "acme/assistant"
description: "A helpful AI assistant"
author: "@acme"
license: "MIT"
tags: ["assistant", "general"]
---

# Agent Purpose
I am a helpful AI assistant that can answer questions and help with tasks...

Vendor Format Mapping Guides

Since OAF is a new specification, we provide mapping guides to help you convert between OAF and existing vendor formats:

Claude Code

Convert to/from Claude Code skills format

Goose

Convert to/from Goose extensions format

Deep Agents

Convert to/from Deep Agents format

Letta

Convert to/from Letta Agent File format

Note: These mapping pages are coming soon. We're working with each vendor to ensure accurate conversion procedures.

Read Spec Raw Markdown

Getting Started

Ready to create your first OAF agent? Here's what you need to know:

  1. Create a directory for your agent (e.g., my-vendor/my-agent/)
  2. Add an AGENTS.md file with metadata and instructions describing the agent's composition and behavior
  3. Optionally add skills, MCP configs, or sub-agents as needed for complex compositions
  4. Deploy to your preferred harness using the mapping guides above

For detailed specifications, examples, and best practices, see the full documentation.

Contributing

The Open Agent Format is a living specification. We welcome feedback, proposals, and contributions from the community. If you'd like to contribute or suggest improvements: