Introduction

Overview

Dhenara Agent DSL (DAD) is an open-source framework built on top of the dhenara-ai Python package. It provides a powerful, expressive, and type-safe domain-specific language (DSL) for defining and executing AI agent workflows. DAD makes it easier to create, compose, and orchestrate AI agents with sophisticated behaviors, while maintaining robust observability and reproducibility.

What is Dhenara Agent DSL?

Dhenara Agent DSL or DAD (available as a Python package named dhenara-agent) is an AI agent framework with a strong focus on:

1. Expressive Agent Definition: Create complex agent workflows using a straightforward, programming language-like approach
2. Component-Based Architecture: Compose reusable components to build sophisticated agent systems
3. Out-of-the-box Support for Multiple LLMs: Switch between different LLM models on the fly
4. Comprehensive Observability: Built-in logging, tracing, and metrics collection for all agent activities using OpenTelemetry and open-source exporters like Zipkin and Jaeger
5. Reproducible Execution: Track and replay agent execution through a run context system, reducing costs by rerunning failed flows without additional AI Model API calls
6. Extensible Node System: Easily create custom node types to extend functionality
7. Resource Management: Flexible management of AI model resources and credentials

Core Concepts

Basic Elements

DAD uses a hierarchical component model that allows for composition and reuse. It is built around three primary types of components:

• Execution Nodes: Atomic execution units that perform specific functions (e.g., making an LLM API call, analyzing a folder, performing file operations like creating/updating files)
• Execution Flows: Collections of nodes or sub-flows with execution logic, supporting sequential execution, conditionals, and loops
• Agents: Higher-level abstractions that can contain flows and other agents, representing complete functional units

Event-Driven Architecture

An event system enables loose coupling between components, allowing agents to react to events, request inputs, and communicate with each other without tight coupling.

Powerful Template Engine

A powerful template engine supports variable substitution, expressions, and hierarchical references, making it easy to build dynamic prompts and process responses.

Execution Model

The execution follows a hierarchical structure:

1. Components (Agents or Flows) define the overall structure
2. Nodes within components perform specific tasks
3. A RunContext manages the execution environment
4. Tracing, logging, and metrics provide visibility into execution

Resource Management

DAD provides a flexible system for managing AI model resources and API credentials, making it easier to work with different LLM providers and models.

Basic Example

Here's a simple example of defining a flow using DAD:

from dhenara.agent.dsl import (
    AIModelNode,
    AIModelNodeSettings,
    FlowDefinition,
    ResourceConfigItem,
)
from dhenara.ai.types import Prompt

# Define a flow
my_flow = FlowDefinition()

# Add an AI model node to the flow
my_flow.node(
    "question_answerer",
    AIModelNode(
        resources=ResourceConfigItem.with_model("claude-3-5-haiku"),
        settings=AIModelNodeSettings(
            system_instructions=["You are a helpful assistant."],
            prompt=Prompt.with_dad_text("Answer the following question: $var{question}"),
        ),
    ),
)

This example defines a simple flow with a single AI model node that uses Claude 3.5 Haiku to answer a question.