AgentML
...Join Community
Open Source • MIT Licensed • Alpha • Building in Public

A Deterministic Language for Agent Behavior

Think of it as HTML for agents. Write once, validate with W3C-compliant schemas, and deploy anywhere with full type safety and traceability. Turn agent design from a prompt craft problem into a system design discipline.

Get StartedView on GitHubDocumentation
$ curl -fsSL sh.agentml.dev | sh
agent.aml
<agentml xmlns="github.com/agentflare-ai/agentml"
       datamodel="ecmascript"
       xmlns:gemini="github.com/agentflare-ai/agentml-go/gemini">

  <datamodel>
    <data id="user_input" expr="''" />
  </datamodel>

  <state id="greeting">
    <onentry>
      <gemini:generate
        model="gemini-2.0-flash-exp"
        location="_event"
        promptexpr="'Welcome user warmly'" />
    </onentry>
    <transition event="action.response" target="classify" />
  </state>

  <state id="classify">
    <transition event="intent.technical" target="technical_help" />
    <transition event="intent.billing" target="billing_help" />
  </state>
</agentml>

Where Structure Meets Intelligence

AgentML combines the usability of XML with the predictability of state machines, giving you complete control over agent behavior.

Readable Structure

XML-like syntax that's intuitive to author, diff, and audit. Human-readable agent definitions that developers can understand at a glance.

Deterministic Logic

State machine core with 100% W3C SCXML conformance ensures every action and transition is governed by declared conditions. No more unpredictable agent behavior.

Schema Validation

Full XSD 1.0 support with automatic import/include resolution. Comprehensive validation ensures your agent definitions are correct before deployment.

Transparent Reasoning

Every step, tool call, and transition is logged as a traceable event. Full replayability for debugging and compliance.

Built for Compliance

Validators and guards ensure actions conform to policy and safety constraints before execution. Identity constraints and type checking built-in.

Schema Caching

Efficient schema reuse with built-in caching for frequently used agent definitions. Optimized for high-throughput production environments.

AMLX Runtime

Powerful runtime that interprets AML documents into live workflows. Combines state machine rigor with modern orchestration.

Thread-Safe Operations

Concurrent agent validation and execution support. Run multiple agents in parallel with confidence and efficient resource management.

MCP Integration

Seamless integration with Model Context Protocol tool servers, APIs, and LLM calls as safe, typed actions.

Coming Soon

We're working on bringing AI-powered AML generation to life. Stay tuned!

AI-Powered Generation

See How Easy It Is

Describe what you want your agent to do, and we'll generate the AML code for you. No complex syntax to learn.

Your generated AML will appear here...
Interactive CLI

Powerful CLI for Every Workflow

Initialize, validate, run, and debug your agents with a comprehensive command-line interface. Built for developer productivity.

  • Compiler-Inspired Validation
    Catch errors before runtime with detailed, actionable error messages
  • W3C SCXML Conformance
    100% conformance with all 193 official W3C SCXML tests
  • Runtime Snapshots
    Debug issues by saving XML snapshots of agent state at each step
agentml-cli

Why AgentML?

The problem with prompt-based agents and why deterministic structure matters

The Prompt Problem

Traditional prompt-based systems rely on inference and natural language, making agent behavior unpredictable and difficult to debug. You can't reason about what an agent will do—you can only hope.

"You are a helpful assistant. Be professional but friendly. Handle customer inquiries..."
→ Unpredictable • Hard to debug • Compliance risk

The AgentML Solution

AgentML makes agent behavior explicit through structured definitions. Every state, transition, and action is declared upfront, giving you the same level of control you have over your code.

<state id="classify">
  <transition event="intent.technical" target="technical_help" />
  <transition event="intent.billing" target="billing_help" />
</state>
→ Predictable • Debuggable • Compliant

System Design Discipline

AgentML transforms agent development from prompt engineering into system design. You define states, transitions, and validations just like you would design any other software system—with precision, clarity, and confidence.

Ready to Build Deterministic Agents?

Join the community building the future of predictable AI agents

Read the DocsStar on GitHub