Agent Definition Language (ADL)

Getting Started

ADL (Agent Definition Language) is an open, declarative, vendor-neutral specification for defining AI agents in a consistent, auditable, and interoperable way. It provides a shared language for describing:

an agent’s identity and purpose
its tools and capabilities
its LLM configuration
its access to knowledge (RAG)
its permissions and sandbox
its dependencies
its governance metadata

If OpenAPI defines APIs, ADL defines agents.

🧠 Why ADL Exists

Enterprises adopting AI agents face several systemic challenges:

Each vendor defines “agents” differently
Tool contracts are inconsistent
RAG pipelines are wired differently across apps
Permissions are rarely explicit
Governance teams have no centralized visibility
Agents are not portable across platforms
Vendor lock-in slows enterprise adoption
Compliance and auditability are fragile or impossible

ADL solves these problems by introducing a single, declarative, versioned artifact that describes what an agent is and what it is allowed to do.

🧩 What ADL Is

ADL defines:

Identity — name, description, role, owner, version
LLM Settings — provider, model, temperature, max tokens
Tools & Actions — typed parameters, descriptions, return schemas
RAG Inputs — indices, types, metadata, paths
Permissions — file I/O, network, env vars
Dependencies — Python packages with optional version pins
Governance — created_by, approved_by, timestamps, version notes

This makes agents:

portable
predictable
auditable
reproducible
interoperable across vendors

🚫 What ADL Is Not

To avoid confusion, ADL explicitly does not define:

❌ A2A (agent-to-agent) communication protocols
❌ runtime tool invocation semantics (e.g., MCP)
❌ prompt templating or formatting
❌ workflow orchestration (Airflow, Temporal, Dagster)
❌ API schemas (OpenAPI already solves that)
❌ message transport (HTTP, gRPC, JSON-RPC)

ADL is laser-focused on definition — not execution.

🔍 ADL vs AI App Definition

ADL is an Agent Definition Language - not a general AI App definition format.

AI apps are broad and may include UI, API layers, deployments, data stores, or business logic.

Agents are specific:

they reason
they call tools
they retrieve knowledge
they act autonomously
they require permission boundaries

ADL models agent competencies, not app-level infrastructure.

This is a key strategic distinction.

🔄 Comparing ADL to Other Standards

ADL vs A2A

A2A: defines how agents communicate
ADL: defines what an agent is

ADL vs MCP

MCP: runtime tool protocol
ADL: declarative definition of tools and capabilities

ADL vs OpenAPI

OpenAPI: describes HTTP services
ADL: describes agent behavior, boundaries, and capabilities

ADL vs Workflow Engines

Workflows = when & how tasks execute
ADL = which agent executes them

🌐 Why Next Moca Open Sourced ADL

Next Moca open-sourced ADL under Apache 2.0 to enable:

✔ Ecosystem-wide interoperability

✔ Enterprise trust and transparency

✔ Neutral governance

✔ Community-driven evolution

✔ Vendor adoption without lock-in

✔ Safe, compliant, standards-based agent deployment

Open sourcing ensures ADL becomes a true standard, not a proprietary configuration format.

📘 Documentation

📄 ADL Spec (JSON Schema) — /schema/adl.schema.json
📚 Examples — /examples/
📐 Roadmap — /ROADMAP.md
🏛 Governance — /GOVERNANCE.md
🤝 Contributing — /CONTRIBUTING.md

🛠 Getting Started

git clone https://github.com/nextmoca/adl.git
cd adl

Validate an ADL Agent File

pip install jsonschema
python tools/validate.py examples/minimal_agent.json

Or using Node:

npm install ajv
node tools/validate.js examples/minimal_agent.json

Minimal Example

{
  "name": "campaign_image_generator",
  "description": "Generate a 1024x1024 marketing image from a creative brief.",
  "role": "Creative Producer",
  "llm": "openai",
  "llm_settings": {
    "temperature": 0,
    "max_tokens": 4096
  },
  "tools": [
    {
      "name": "generate_campaign_image",
      "description": "Generate a high-quality image from a prompt.",
      "parameters": [
        {
          "name": "prompt",
          "type": "string",
          "description": "Image prompt",
          "required": true
        }
      ],
      "invocation": { "type": "python_function" }
    }
  ],
  "rag": []
}

Tags: ai language dsl

Last modified 15 March 2026