Agents and Jobs

Learn how AI agents work together through jobs in Atlas

Agents and Jobs

Agents are the workers in Atlas, and jobs define how they work together. Let’s explore both concepts and how they interact.

Agents: Your AI Workers

Agents are autonomous AI entities that perform specific tasks. Think of them as specialized team members, each with their own expertise.

Types of Agents

1. LLM Agents 🤖

• Powered by language models (Claude, GPT-4, etc.)
• Handle natural language tasks
• Can use tools and reason through problems

2. System Agents ⚙️

• Built-in agents provided by Atlas
• Examples: ConversationAgent, MemoryAgent
• Optimized for specific platform tasks

3. Remote Agents 🌐

• External agents accessed via protocols (ACP, REST)
• Integrate third-party AI services
• Enable distributed architectures

Defining an Agent

Here’s a simple LLM agent:

agents:
  researcher:
    type: "llm"
    model: "claude-3-5-sonnet-20241022"
    purpose: "Research and analyze information"
    prompts:
      system: |
        You are a research specialist. Your role is to:
        - Find accurate, relevant information
        - Analyze data and identify patterns
        - Provide clear, evidence-based insights
        - Cite sources when possible
    config:
      temperature: 0.3  # Lower = more focused
      max_tokens: 2000

Jobs: Orchestrating Agent Work

Jobs define workflows - how agents work together to accomplish tasks. They’re triggered by signals and executed by agents.

Job Structure

jobs:
  analyze-data:
    name: "Data Analysis Pipeline"
    description: "Research, analyze, and report on data"
    
    triggers:
      - signal: "analyze-request"
        condition:
          # Only run for CSV files
          - type: "json_logic"
            logic:
              "==": [{"var": "fileType"}, "csv"]
    
    execution:
      strategy: "sequential"  # or "parallel"
      agents:
        - id: "researcher"
          input_source: "signal"
          task: "Analyze the provided data"
        
        - id: "reporter"
          input_source: "previous"
          task: "Create a summary report"

Execution Strategies

Sequential 📝

• Agents run one after another
• Each agent can use the previous agent’s output
• Good for: pipelines, step-by-step processes

Parallel ⚡

• Agents run simultaneously
• Faster but no data sharing between agents
• Good for: independent tasks, gathering multiple perspectives

Input Sources

Agents can receive input from:

• signal - The triggering signal’s data
• previous - The previous agent’s output
• context - Shared session context
• combined - Combination of sources

How Agents and Jobs Work Together

Here’s the flow:

1. Signal triggers job → “analyze-request” signal received
2. Job starts session → SessionSupervisor created
3. Agents activated → Each agent in the job is prepared
4. Execution begins → Agents work according to strategy
5. Results collected → Outputs gathered and returned

Practical Example: Customer Support

Let’s build a customer support system:

agents:
  # Understands customer intent
  intent-analyzer:
    type: "llm"
    model: "claude-3-5-haiku-20241022"
    purpose: "Analyze customer support requests"
    prompts:
      system: "Categorize support requests and extract key information"
  
  # Provides solutions
  solution-provider:
    type: "llm"
    model: "claude-3-5-sonnet-20241022"
    purpose: "Provide detailed solutions"
    prompts:
      system: "You are a technical support expert..."
  
  # Handles conversations
  support-chat:
    type: "system"
    agent: "conversation"
    config:
      model: "claude-3-5-sonnet-20241022"
      system_prompt: "You are a friendly support agent..."

jobs:
  # Automated ticket handling
  handle-ticket:
    triggers:
      - signal: "new-ticket"
    execution:
      strategy: "sequential"
      agents:
        - id: "intent-analyzer"
          input_source: "signal"
        - id: "solution-provider"
          input_source: "previous"
  
  # Interactive support
  live-support:
    triggers:
      - signal: "start-chat"
    execution:
      strategy: "single"
      agents:
        - id: "support-chat"
          input_source: "signal"

Agent Capabilities

Tools

Agents can use tools to extend their capabilities:

agents:
  researcher:
    type: "llm"
    tools: ["web-search", "calculator", "file-reader"]

Memory

Agents can access workspace memory:

agents:
  assistant:
    type: "llm"
    config:
      memory_enabled: true
      context_window: 10  # Recent memories

Best Practices

1. Single Responsibility

Each agent should have one clear purpose. Create multiple specialized agents rather than one do-everything agent.

2. Right Model for the Task

• Use smaller, faster models (Haiku) for simple tasks
• Use larger models (Sonnet, Opus) for complex reasoning
• Use system agents for platform operations

3. Clear Task Definitions

When agents work in sequence, provide clear task instructions:

agents:
  - id: "analyzer"
    task: "Extract the top 3 issues from the data"
  - id: "solver"
    task: "Propose solutions for each issue identified"

4. Error Handling

Plan for failures:

jobs:
  robust-pipeline:
    config:
      timeout: "300s"
      retry_on_failure: true
      max_retries: 3

Next Steps

• Learn about Signals that trigger jobs
• Explore Interactive Mode
• Try the Multi-Agent Tutorial