Sherpa AI Concepts

This page introduces the key concepts and terminology used throughout the Sherpa AI framework. Understanding these concepts is essential for effectively using and extending the framework.

Core Components

Agents

Agents are the primary actors in Sherpa AI. They are specialized AI components designed to perform specific tasks or provide expertise in particular domains. Each agent:

• Encapsulates a language model (LLM)

• Has a defined role and description

• Maintains its own belief state

• Can access shared memory

• Follows a policy for decision making

• Can execute a set of actions

Examples include the MLEngineer agent for machine learning tasks and the QAAgent for question-answering tasks.

Policies

Policies govern how agents make decisions. They implement the reasoning patterns and strategies that agents use to select actions. Sherpa AI includes several policy types:

• ReactPolicy: Implements the ReAct (Reasoning + Acting) framework

• ReactStateMachinePolicy: Extends ReAct with state machine capabilities

• ChatStateMachinePolicy: Implements chat-based decision making with state management

• AgentFeedbackPolicy: Enables decision making based on feedback from other agents

Each policy type provides different approaches to:

• Action selection

• Context processing

• Response generation

• State management

Actions

Actions are concrete operations that agents can perform. They represent the capabilities available to agents and include:

• Information retrieval (e.g., GoogleSearch, ArxivSearch)

• Output synthesis

• Deliberation

• External API interactions

Each action has:

• A unique name

• Required arguments

• Usage description

• Optional belief state access

• Input validation logic

Memory Systems

Belief

The Belief system represents an agent’s understanding of its current state and context. It includes:

• Current task information

• Action history

• State machine status (if applicable)

• Internal event log

• Context management

Shared Memory

Shared Memory enables communication and information sharing between agents. It provides:

• Cross-agent data persistence

• Observation mechanisms

• Result storage

• Context sharing

State Management

State Machines

State machines provide structured workflow management for agents. They:

• Define possible states

• Control state transitions

• Include state-specific behaviors

• Guide action selection

Event System

The event system manages the flow of information and state changes:

• Tracks action execution

• Records agent decisions

• Maintains history

• Enables feedback loops

Integration Components

Models

Models represent the underlying language models (LLMs) that power agent capabilities. The framework:

• Supports multiple LLM providers

• Handles response processing

• Manages token usage

• Provides standardized interfaces

Prompts

The prompt system manages structured inputs to language models:

• Template-based generation

• Variable substitution

• Version control

• Context formatting

Output Processing

Output processors handle and validate agent responses:

• Citation validation

• Response formatting

• Error handling

• Quality checks

Best Practices

When working with Sherpa AI:

1. Agent Design: * Define clear agent responsibilities * Choose appropriate policies * Configure relevant actions

2. Memory Management: * Use shared memory for cross-agent communication * Maintain clean belief states * Handle context appropriately

3. Action Implementation: * Validate inputs thoroughly * Handle errors gracefully * Document usage clearly

4. Policy Selection: * Match policy to task requirements * Configure appropriate response formats * Handle state transitions carefully

For detailed implementation examples and API references, see the API Documentation.