Understanding Agents

Agents are at the core of the BeeAI platform. They represent autonomous components that can perform specific tasks, process information, or interact with users. This document explains agent concepts, their structure, and how they function within the BeeAI ecosystem.

What are Agents?

In BeeAI, an agent is a self-contained module that:

Implements the Agent Communication Protocol (ACP)
Has clearly defined inputs and outputs
Performs a specific function or set of related functions
Can operate independently or as part of a composition
Is packaged as a Docker container

Conceptually, agents can be thought of as specialized “workers” that excel at specific tasks, from simple operations like text summarization to complex workflows like research, content generation, or data analysis.

Agent Architecture

Every agent in BeeAI has these common components:

1. Agent Interface

The interface defines how the agent communicates with the outside world through:

Input Schema: Defines the expected input format
Output Schema: Specifies the structure of the agent’s response
Metadata: Information about the agent’s capabilities, requirements, and usage

2. Agent Logic

The core functionality of the agent, which can include:

Processing input data
Making API calls to external services
Using language models for text generation
Utilizing specialized tools (search, weather, etc.)
Performing computations or data analysis

3. Container Runtime

All agents run in containerized environments (Docker), ensuring:

Consistent execution across different platforms
Isolation from the host system and other agents
Dependency management
Scalability and reproducibility

Agent Types

BeeAI supports various types of agents:

By Interaction Pattern

Chat Agents: Conversational agents designed for back-and-forth interaction with users
Hands-Off Agents: Single-purpose agents that process input and produce output without intermediate steps
Custom UI Agents: Agents with specialized interfaces for specific use cases

By Framework

BeeAI Framework Agents: Built using the official BeeAI Framework
Community Framework Agents: Built with frameworks like LangChain, CrewAI, LlamaIndex, etc.
Custom Agents: An agent implementation that implements ACP protocol built with potentialy any framework

By Purpose

Tool Agents: Perform specific operations like search, data processing, or code generation
Assistant Agents: Provide conversational help and information to users
Composer Agents: Coordinate the execution of multiple agents (like sequential_workflow)
Domain-Specific Agents: Specialized for particular fields like legal, medical, or financial tasks

Agent Metadata

Agent metadata provides essential information about an agent, including:

Name: Unique identifier for the agent
Description: Brief summary of the agent’s purpose
Documentation: Detailed explanation of how the agent works
Framework: The framework or technology used to build the agent
Examples: Sample usage and expected outputs
Environment Variables: Configuration options for the agent
UI Type: How the agent should be presented in the user interface
Links: References to source code, documentation, etc.

Here’s an example of agent metadata:

name: chat
description: AI-powered conversational system with memory and tools
metadata:
  programming_language: Python
  license: Apache 2.0
  framework: BeeAI
  documentation: |
    The agent is an AI-powered conversational system designed to process user messages,
    maintain context, and generate intelligent responses...
  ui:
    type: chat
    user_greeting: How can I help you?
  examples:
    cli:
      - command: 'beeai run chat "What is the weather like in Paris?"'
        name: With tools
        description: Run agent with tools.
        output: The current temperature in Paris is 12°C with partly cloudy skies.
  env:
    - name: LLM_MODEL
      description: Model to use from the specified OpenAI-compatible API.
    - name: LLM_API_BASE
      description: Base URL for OpenAI-compatible API endpoint
    - name: LLM_API_KEY
      description: API key for OpenAI-compatible API endpoint

Standardized Agent Interfaces

BeeAI provides standardized UI patterns for agent interaction:

Chat Interface

Designed for conversational agents:

Full conversation history and context
Message-by-message interaction
Streaming responses
Suitable for assistants and complex interaction flows

Hands-Off Interface

For single-input, single-output agents:

One input field for the user request
Processes the request and returns a complete response
No intermediate interaction
Suitable for transformation, generation, or analysis tasks

Custom Interfaces

For specialized agent requirements:

JSON input interface for complex data structures
Custom visualizations
Application-specific UI elements

Agent Lifecycle

The lifecycle of a BeeAI agent includes these stages:

Discovery: Agents are found in the registry or from custom sources
Installation: Agents are downloaded and installed locally
Configuration: Required environment variables are set
Execution: The agent runs in response to user requests
Updating: Agents are updated to newer versions as needed
Removal: Agents can be uninstalled when no longer needed

Agent Management

BeeAI provides tools for managing agents:

Listing Agents: beeai list shows all available agents
Installing: beeai install <name> installs an agent
Running: beeai run <name> [input] executes an agent
Inspecting: beeai info <name> shows agent details
Viewing Logs: beeai logs <name> displays agent logs
Removing: beeai remove <name> uninstalls an agent

Creating Agents

To create your own agent:

Choose a development framework (BeeAI Framework, LangChain, etc.)
Implement the agent logic following the ACP standard
Package the agent as a Docker container
Test the agent locally
Share the agent with others

See Building Agents for detailed instructions.

Next Steps

description: “Understand AI agents in BeeAI: types, capabilities, lifecycle management, and how they work across different frameworks through the ACP protocol.”

Agents

In BeeAI, agents are self-contained AI programs that can perform specific tasks or engage in conversations. What makes BeeAI unique is that agents from any framework - CrewAI, LangChain, AutoGen, or custom implementations - all work seamlessly together through the Agent Communication Protocol (ACP).

What Are Agents?

An agent in BeeAI is an AI program that:

Accepts input in natural language or structured formats
Processes requests using AI models and custom logic
Produces output such as text, files, or structured data
Runs independently in its own containerized environment
Communicates through the standardized ACP protocol

Agent vs. Traditional Software

Traditional Software	BeeAI Agents
Fixed, programmed responses	Dynamic, AI-powered responses
Rigid input/output formats	Natural language interaction
Single-purpose functions	Adaptable task handling
Manual integration required	Universal ACP compatibility

Agent Types by Interface

Chat Agents

Interactive agents that maintain conversation context and support back-and-forth dialogue.

Characteristics:

Persistent conversation memory
Interactive user interface
Real-time streaming responses
Context-aware responses

Example Usage:

beeai run chat
>>> Hello, how can you help me?
I'm an AI assistant ready to help with questions, analysis, writing, and more...
>>> Can you explain quantum computing?
Certainly! Quantum computing is a revolutionary approach to computation...
>>> How does it differ from classical computing?
Great follow-up question! The key differences are...

Use Cases:

General assistance and Q&A
Interactive problem-solving
Educational conversations
Customer support scenarios

Hands-off Agents

Task-focused agents that take a single instruction and work autonomously to complete it.

Characteristics:

Single input, complete output
Autonomous task execution
Progress indicators during work
Comprehensive final results

Example Usage:

beeai run research-assistant "Research the latest developments in renewable energy"
🔍 Gathering information from sources...
📊 Analyzing data patterns...
📝 Generating comprehensive report...
✅ Research complete! Generated files: renewable_energy_report.pdf

Use Cases:

Research and analysis tasks
Content generation
Code review and analysis
Report creation

Workflow Agents

Meta-agents that orchestrate multiple other agents to complete complex, multi-step tasks.

Characteristics:

Orchestrates multiple agents
Manages data flow between steps
Handles different completion times
Combines outputs intelligently

Example Usage:

beeai run sequential-workflow
# Interactive setup:
# Step 1: research-assistant "Research AI trends"
# Step 2: document-creator "Create executive summary"
# Step 3: presentation-maker "Create slides"

Use Cases:

Complex business processes
Multi-stage content creation
Data processing pipelines
Automated reporting workflows

Agent Types by Framework

CrewAI Agents

Agents built using the CrewAI framework, designed for collaborative multi-agent workflows.

Example Agent:

from crewai import Agent, Task, Crew
from acp_sdk.server import create_server

def run_crew_agent(input_text):
    researcher = Agent(
        role='Senior Research Analyst',
        goal='Uncover cutting-edge developments',
        backstory='Expert in analyzing complex topics'
    )
    
    task = Task(
        description=input_text,
        agent=researcher
    )
    
    crew = Crew(agents=[researcher], tasks=[task])
    return crew.kickoff()

server = create_server(
    name="crewai-researcher",
    description="Research agent built with CrewAI",
    agent_callable=run_crew_agent
)

Strengths:

Multi-agent collaboration
Role-based agent design
Built-in task management
Hierarchical workflows

LangChain Agents

Agents leveraging LangChain’s extensive ecosystem of tools and integrations.

Example Agent:

from langchain.agents import create_react_agent
from langchain_openai import ChatOpenAI
from langchain.tools import DuckDuckGoSearchRun
from acp_sdk.server import create_server

def run_langchain_agent(input_text):
    llm = ChatOpenAI()
    tools = [DuckDuckGoSearchRun()]
    agent = create_react_agent(llm, tools, prompt)
    return agent.invoke({"input": input_text})

server = create_server(
    name="langchain-researcher",
    description="Research agent with web search",
    agent_callable=run_langchain_agent
)

Strengths:

Extensive tool ecosystem
Memory management
Chain composition
Vector database integration

AutoGen Agents

Agents built with Microsoft’s AutoGen for multi-agent conversations.

Example Agent:

import autogen
from acp_sdk.server import create_server

def run_autogen_agent(input_text):
    config_list = [{"model": "gpt-4", "api_key": "..."}]
    
    assistant = autogen.AssistantAgent(
        name="assistant",
        llm_config={"config_list": config_list}
    )
    
    user_proxy = autogen.UserProxyAgent(
        name="user_proxy",
        human_input_mode="NEVER"
    )
    
    user_proxy.initiate_chat(assistant, message=input_text)
    return assistant.last_message()["content"]

server = create_server(
    name="autogen-assistant",
    description="Multi-agent conversation with AutoGen",
    agent_callable=run_autogen_agent
)

Strengths:

Multi-agent conversations
Human-in-the-loop capabilities
Code execution environment
Group chat coordination

Custom Framework Agents

Agents built with custom frameworks or from scratch, tailored for specific use cases.

Example Agent:

from acp_sdk.server import create_server
import openai

class CustomAnalysisAgent:
    def __init__(self):
        self.client = openai.OpenAI()
        self.analysis_prompt = """
        You are a data analysis expert. Analyze the provided data
        and generate insights with visualizations.
        """
    
    def process(self, input_text):
        response = self.client.chat.completions.create(
            model="gpt-4",
            messages=[
                {"role": "system", "content": self.analysis_prompt},
                {"role": "user", "content": input_text}
            ]
        )
        return response.choices[0].message.content

agent = CustomAnalysisAgent()

server = create_server(
    name="custom-analyzer",
    description="Custom data analysis agent",
    agent_callable=agent.process
)

Strengths:

Complete control over behavior
Optimized for specific tasks
Minimal dependencies
Custom integration capabilities

Agent Lifecycle

Discovery Phase

Agents become available through various discovery mechanisms:

Sources:

Official Registry: Curated agents from BeeAI team
Community Registry: Community-contributed agents
Docker Registries: Public and private container registries
GitHub Repositories: Source code repositories with agent definitions
Custom Registries: Enterprise and organizational agent repositories

Installation and Deployment

When an agent is first requested, BeeAI handles the complete deployment process:

Resource Management

Automatic Scaling:

Scale Up: Agents start when first requested
Scale Out: Multiple instances for high demand
Scale Down: Automatic shutdown after idle period
Resource Limits: CPU, memory, and storage quotas

Health Monitoring:

Liveness Probes: Detect unresponsive agents
Readiness Probes: Ensure agents are ready for requests
Performance Metrics: Response times and success rates
Automatic Recovery: Restart failed agents

Agent Configuration

Environment Variables

Agents require configuration for external services and behavior:

Common Variables:

# LLM Configuration
LLM_API_KEY=your-api-key-here
LLM_MODEL=gpt-4
LLM_API_BASE=https://api.openai.com/v1

# Agent-Specific Configuration
AGENT_TIMEOUT=300
AGENT_MAX_TOKENS=2000
AGENT_TEMPERATURE=0.7

# Platform Integration
OTEL_EXPORTER_OTLP_ENDPOINT=http://collector:8335
PLATFORM_URL=http://beeai-platform:8333

Security:

Variables stored encrypted in Kubernetes secrets
Automatic injection into agent containers
No hardcoded credentials in agent images
Support for external secret management systems

Agent Metadata

Every agent includes comprehensive metadata describing its capabilities:

# agent.yaml
name: research-assistant
description: "AI agent that conducts research and analysis on any topic"
version: "2.1.0"
author: "BeeAI Team <team@beeai.dev>"
license: "Apache-2.0"
framework: "crewai"

# Categorization
category: "research"
tags: ["research", "analysis", "web-search", "reports"]
difficulty: "beginner"

# Resource requirements
resources:
  cpu: "500m"
  memory: "1Gi"
  storage: "10Gi"

# Environment configuration
environment:
  - name: LLM_API_KEY
    description: "API key for language model access"
    required: true
  - name: SEARCH_API_KEY
    description: "API key for web search functionality"
    required: false
  - name: MAX_SOURCES
    description: "Maximum number of sources to research"
    required: false
    default: "10"

# User interface configuration
ui:
  type: "hands-off"
  user_greeting: "What topic would you like me to research?"
  expected_input: "A research topic or question"
  output_format: "Comprehensive research report with sources"

# Usage examples
examples:
  cli:
    - name: "Technology Research"
      command: 'beeai run research-assistant "Latest developments in quantum computing"'
      description: "Research current quantum computing breakthroughs"
      expected_output: "Detailed report with recent developments, key players, and future outlook"
      
    - name: "Market Analysis"
      command: 'beeai run research-assistant "Electric vehicle market trends 2024"'
      description: "Analyze current EV market conditions"
      expected_output: "Market analysis with statistics, trends, and forecasts"

# Documentation
documentation: |
  # Research Assistant Agent
  
  This agent conducts comprehensive research on any topic using web search,
  academic sources, and AI analysis to produce detailed reports.
  
  ## Capabilities
  - Web search and source gathering
  - Content analysis and synthesis
  - Report generation with citations
  - Trend identification and analysis
  
  ## Best Practices
  - Be specific with research topics for best results
  - Include time frames when relevant (e.g., "2024 trends")
  - Ask for specific aspects if you want focused research

Agent Communication

Input Formats

Agents accept various input formats through the ACP protocol:

Simple Text:

{
  "input": [
    {
      "parts": [
        {
          "content": "Research renewable energy trends",
          "role": "user"
        }
      ]
    }
  ]
}

Structured Input:

{
  "input": [
    {
      "parts": [
        {
          "content": "Analyze the attached dataset",
          "role": "user"
        },
        {
          "content": "base64-encoded-csv-data",
          "content_type": "text/csv",
          "role": "user"
        }
      ]
    }
  ]
}

Multi-modal Input:

{
  "input": [
    {
      "parts": [
        {
          "content": "Describe this image and suggest improvements",
          "role": "user"
        },
        {
          "content": "base64-encoded-image-data",
          "content_type": "image/png",
          "role": "user"
        }
      ]
    }
  ]
}

Output Formats

Agents produce various types of output:

Text Responses:

{
  "type": "message_part",
  "content": "Based on my research, renewable energy adoption..."
}

File Artifacts:

{
  "type": "artifact",
  "name": "research_report.pdf",
  "content": "base64-encoded-pdf-content",
  "content_type": "application/pdf",
  "content_encoding": "base64"
}

Progress Updates:

{
  "type": "generic",
  "agent_name": "research-assistant",
  "status": "gathering_sources",
  "progress": 0.3,
  "message": "Found 15 relevant sources, analyzing content..."
}

Agent Development Best Practices

Design Principles

Single Responsibility:

Each agent should have a clear, focused purpose
Avoid creating “do-everything” agents
Compose complex workflows from specialized agents

User-Centric Design:

Clear, helpful descriptions and examples
Intuitive input expectations
Meaningful progress indicators
Comprehensive error messages

Reliability:

Graceful error handling
Appropriate timeouts
Resource usage optimization
Consistent output quality

Performance Optimization

Efficient Resource Usage:

# Good: Stream responses for long-running tasks
async def long_running_agent(input_text):
    yield MessagePartEvent("Starting analysis...")
    
    for step in analysis_steps:
        result = await process_step(step)
        yield MessagePartEvent(f"Completed {step}: {result}")
    
    yield MessagePartEvent("Analysis complete!")

# Bad: Return everything at once after long processing
def blocking_agent(input_text):
    # Long processing with no feedback
    result = expensive_operation(input_text)
    return result  # User waits with no progress indication

Memory Management:

# Good: Process data in chunks
def process_large_dataset(data):
    for chunk in chunk_data(data, size=1000):
        yield process_chunk(chunk)

# Bad: Load everything into memory
def process_all_at_once(data):
    return [process_item(item) for item in data]  # Memory intensive

Error Handling

User-Friendly Errors:

try:
    result = call_external_api()
except APIKeyError:
    raise ValueError(
        "API key is missing or invalid. "
        "Please check your LLM_API_KEY environment variable."
    )
except RateLimitError:
    raise ValueError(
        "Rate limit exceeded. Please try again in a few minutes."
    )
except NetworkError:
    raise ValueError(
        "Network connection failed. Please check your internet connection."
    )

Agent Security

Input Validation

Sanitize Inputs:

def validate_input(input_text):
    # Check for malicious content
    if contains_injection_patterns(input_text):
        raise ValueError("Input contains potentially malicious content")
    
    # Limit input size
    if len(input_text) > MAX_INPUT_SIZE:
        raise ValueError(f"Input too large. Maximum size: {MAX_INPUT_SIZE}")
    
    return sanitize_text(input_text)

Secret Management

Never Hardcode Secrets:

# Good: Use environment variables
api_key = os.getenv("LLM_API_KEY")
if not api_key:
    raise ValueError("LLM_API_KEY environment variable required")

# Bad: Hardcoded secrets
api_key = "sk-hardcoded-key-here"  # Security risk!

Container Security

Run as Non-Root User:

# Create non-root user
RUN useradd -m -u 1000 agent

# Switch to non-root user
USER agent

# Set secure permissions
RUN chmod 755 /app

Agent Monitoring and Observability

Health Checks

Every agent implements health endpoints:

@app.get("/ping")
async def health_check():
    return {
        "status": "healthy",
        "version": "1.0.0",
        "timestamp": datetime.utcnow().isoformat(),
        "dependencies": {
            "llm_provider": await check_llm_connection(),
            "database": await check_database_connection()
        }
    }

Metrics Collection

Performance Metrics:

Request/response times
Success/failure rates
Resource utilization
Concurrent request handling

Business Metrics:

Popular agent features
User satisfaction indicators
Task completion rates
Error patterns and causes

Logging

Structured Logging:

import structlog

logger = structlog.get_logger()

def process_request(input_text):
    logger.info(
        "processing_request",
        input_length=len(input_text),
        agent_version="1.0.0"
    )
    
    try:
        result = process(input_text)
        logger.info(
            "request_completed",
            output_length=len(result),
            processing_time=elapsed_time
        )
        return result
    except Exception as e:
        logger.error(
            "request_failed",
            error_type=type(e).__name__,
            error_message=str(e)
        )
        raise

Future Agent Capabilities

Planned Enhancements

Multi-Agent Collaboration:

Direct agent-to-agent communication
Shared working memory
Coordinated task execution
Conflict resolution mechanisms

Advanced Interfaces:

Voice input/output support
Real-time collaboration features
Mobile-optimized interfaces
AR/VR integration capabilities

Enhanced AI Capabilities:

Multi-modal processing (text, image, audio, video)
Long-term memory and learning
Personalization and adaptation
Reasoning chain visualization

Enterprise Features:

Advanced access controls
Audit trails and compliance
Custom deployment policies
Integration with enterprise systems

Agents in BeeAI represent a new paradigm where AI capabilities from any framework can be shared, discovered, and orchestrated through standardized interfaces, creating a rich ecosystem of specialized AI tools that work seamlessly together.

Introduction

How To

Core Concepts

Deployment

Community and Support

​Understanding Agents

​What are Agents?

​Agent Architecture

​1. Agent Interface

​2. Agent Logic

​3. Container Runtime

​Agent Types

​By Interaction Pattern

​By Framework

​By Purpose

​Agent Metadata

​Standardized Agent Interfaces

​Chat Interface

​Hands-Off Interface

​Custom Interfaces

​Agent Lifecycle

​Agent Management

​Creating Agents

​Next Steps

​description: “Understand AI agents in BeeAI: types, capabilities, lifecycle management, and how they work across different frameworks through the ACP protocol.”

​Agents

​What Are Agents?

​Agent vs. Traditional Software

​Agent Types by Interface

​Chat Agents

​Hands-off Agents

​Workflow Agents

​Agent Types by Framework

​CrewAI Agents

​LangChain Agents

​AutoGen Agents

​Custom Framework Agents

​Agent Lifecycle

​Discovery Phase

​Installation and Deployment

​Resource Management

​Agent Configuration

​Environment Variables

​Agent Metadata

​Agent Communication

​Input Formats

​Output Formats

​Agent Development Best Practices

​Design Principles

​Performance Optimization

​Error Handling

​Agent Security

​Input Validation

​Secret Management

​Container Security

​Agent Monitoring and Observability

​Health Checks

​Metrics Collection

​Logging

​Future Agent Capabilities

​Planned Enhancements

Understanding Agents

What are Agents?

Agent Architecture

1. Agent Interface

2. Agent Logic

3. Container Runtime

Agent Types

By Interaction Pattern

By Framework

By Purpose

Agent Metadata

Standardized Agent Interfaces

Chat Interface

Hands-Off Interface

Custom Interfaces

Agent Lifecycle

Agent Management

Creating Agents

Next Steps

description: “Understand AI agents in BeeAI: types, capabilities, lifecycle management, and how they work across different frameworks through the ACP protocol.”

Agents

What Are Agents?

Agent vs. Traditional Software

Agent Types by Interface

Chat Agents

Hands-off Agents

Workflow Agents

Agent Types by Framework

CrewAI Agents

LangChain Agents

AutoGen Agents

Custom Framework Agents

Agent Lifecycle

Discovery Phase

Installation and Deployment

Resource Management

Agent Configuration

Environment Variables

Agent Metadata

Agent Communication

Input Formats

Output Formats

Agent Development Best Practices

Design Principles

Performance Optimization

Error Handling

Agent Security

Input Validation

Secret Management

Container Security

Agent Monitoring and Observability

Health Checks

Metrics Collection

Logging

Future Agent Capabilities

Planned Enhancements