Agent Applications

Agent Applications is a markdown-first packaging format for applications that an agent can inspect and operate through a CLI.

The core idea is simple:

• APP.md defines the application contract
• app/ contains the runnable payload and owned state
• skills/ contains local agent guidance
• schemas/ is optional for richer validation and machine-readable shapes

Repository contents

• SPEC.md - the current Agent Applications specification
• POSSIBLE-IMPROVEMENTS.md - non-normative ideas for tightening interoperability in a future revision
• skills/agentapplications/SKILL.md - repo-local skill for inspecting and safely using Agent Applications packages
• example/agentic-to-do/ - reference example of an Agent Application package

Why this exists

Agent Applications aims to give agents a usable middle layer between raw code and opaque tool integrations:

• human-readable package contracts
• application-owned state instead of prompt-memory state
• JSON-first CLI interaction
• local skills for safe operation

Example package shape

app-package/
|-- APP.md
|-- app/
|-- skills/
|   `-- <skill>/SKILL.md
`-- schemas/

Example app

The repository includes example/agentic-to-do, a minimal persistent to-do app that demonstrates:

• an APP.md manifest
• JSON-first command execution
• application-owned state
• local package skills
• behavioral contract tests

Status

This repo is an active spec-and-example workspace. SPEC.md is the source of truth for the current format. POSSIBLE-IMPROVEMENTS.md is a non-normative queue of ideas that may inform a later revision.