AML — Agent Markup Language

A lightweight, XML-inspired markup language that enables agents to declaratively invoke modular, scoped, and nestable skills directly inside prompts.

Overview

AML provides two independent layers:

• Runtime Layer: A high-performance parser and executor implemented in Rust + PyO3
• Distribution & Knowledge Layer: An APM package for agent tooling ecosystems

Quick Start

pip install aml

from aml import parse, execute, AmlRegistry

doc = parse("""
<skill interface="code-review" language="python">
  <param name="file">src/auth.py</param>
  Review this module for security issues.
</skill>
""")

registry = AmlRegistry()
registry.register_interface("code-review", "Review code for issues")
registry.register_implementation(
    "python-review", "code-review",
    language="python"
)

result = execute(doc, registry)

AML Syntax

<!-- Invoke a skill -->
<skill interface="testing" language="python" retries="2">
  <param name="target">src/</param>
  Run all tests with coverage.
</skill>

<!-- Define an interface -->
<skill define="interface" name="testing">
  Execute automated tests and report results.
</skill>

<!-- Define an implementation -->
<skill define="implementation" name="pytest-runner" implements="testing" language="python">
  Run tests using pytest with coverage reporting.
</skill>

Key Features

• Declarative: Skills are invoked via markup, not imperative code
• Composable: Full nesting support with bottom-up execution
• Scoped: Content inside tags is the skill's operating scope
• Resolvable: Interface/implementation model with hint-based resolution
• Safe: Results are escaped; no re-parsing attacks

Documentation

• Language Specification
• Conformance Suite
• API Reference (Rust)
• API Reference (Python)

Architecture

┌─────────────────────────────────────────┐
│  Agent Prompt (contains AML tags)        │
└─────────────┬───────────────────────────┘
              │ parse()
              ▼
┌─────────────────────────────────────────┐
│  AST (Document → Nodes)                  │
└─────────────┬───────────────────────────┘
              │ resolve() + execute()
              ▼
┌─────────────────────────────────────────┐
│  SkillRegistry → Resolver → Executor     │
└─────────────┬───────────────────────────┘
              │ result injection
              ▼
┌─────────────────────────────────────────┐
│  Output (AML tags replaced by results)   │
└─────────────────────────────────────────┘

Licence

Apache 2.0 — see LICENSE for details.