AI Agents

What It Is

• A system that observes its environment, makes decisions, uses tools, and acts toward a goal.

Why It Is Needed

• Limitations of a standalone LLM call
  • It generally generates answers using the context already provided
  • Text generation alone cannot perform actions in external systems
• Limitations of predefined workflows
  • Developers must define the execution path in advance
  • Hard to handle unexpected situations
• What agents add
  • Dynamic decision-making
  • Use tools to retrieve information and perform actions
  • Adapt subsequent actions based on observed results
  • Autonomously complete multi-step tasks

How to Implement It

• Agentic Design Patterns
  • ReAct
    • How to Implement It
    • What It Is
      • A pattern that alternates reasoning and acting
      • An agentic pattern that alternates model reasoning with external actions, rather than relying solely on a single internal plan or answer.
    • How it works
      • Reasoning
        • Decide the next step based on the task and current state
      • Action
        • Take an action or invoke a tool
      • Observation
        • Observe the tool result
      • Reasoning
        • Revise the next steps using newly observed information
      • Repeat until the goal is reached or a stopping condition is met
  • Other
    • Andrew NG
      • Four Agentic Design Patterns
        • Reflection
          • Evaluation findings
          • Revision of results or plans based on feedback
        • Planning
          • Disaggregation of tasks into steps
          • Adjustment of plans to results of implementation
        • Tool Use (tool use)
          • Let the model call external tools for information or perform operations
        • Multi-Agent Corporation (Multi-Intelligence Collaboration)
          • Get multiple intelligents to do their job.
      • Different from React.
        • Multi-Agent
    • Other
      • 
    • Prompt Chaining
      • Core Mechanisms
        • Multiple LLM calls executed in a fixed order
        • Analogies in Traditional Engineering
          • Pipeline
        • Use Cases
          • Translation
          • Content treatment of water currents
    • Routing
      • Core Mechanisms
        • Selecting different processing branches according to mandate
      • Analogies in Traditional Engineering
        • Router
        • Policy Mode
      • Use Cases
        • Select a model or Agent by task type
    • Parallelization
      • Core Mechanisms
        • Separate sub-missions handled in parallel and aggregated
      • Analogies in Traditional Engineering
        • Fan-out/Fan-in
        • MapReduce
      • Use Cases
        • Parallel analysis
        • Code Review
    • Planner–Executor
      • Core Mechanisms
        • Mr. Plan, implement and adjust.
      • Analogies in Traditional Engineering
        • Workstream engine
        • Task Scheduler
      • Use Cases
        • Complex multi-step tasks
    • Evaluator–Optimizer
      • Core Mechanisms
        • Generate → Rating → Modify → Rescoring
      • Analogies in Traditional Engineering
        • Test feedback loop
      • Use Cases
        • Code Generation
        • Content Optimization
    • Orchestrator–Workers
      • Core Mechanisms
        • Main Agent dynamic decomposition task assigned to Agent
      • Analogies in Traditional Engineering
        • Master–Worker
      • Use Cases
        • Complex multidisciplinary tasks
    • Multi-Agent Collaboration
      • Core Mechanisms
        • Multi-Agent consultations, interaction, joint decision-making
      • Analogies in Traditional Engineering
        • Actor
        • Distributive collaboration
      • Use Cases
        • Multi-role discussion
        • Joint research
    • Human-in-the-Loop
      • Core Mechanisms
        • Introduce manual confirmation in key decisions or actions
      • Analogies in Traditional Engineering
        • Manual approval of workflows
      • Use Cases
        • High-risk operations
        • Production Publication
• Core Components
  • Environment
    • What It Is
      • Intelligence is in the environment.
    • Environmental feedback
      • Observation
        • Observations returned by environment or tool
  • Tools
    • Classification
      • Knowledge enhancement
        • What It Is
          • Retrieve External Information
          • Search, SQL Query, Document Search
      • Capacity development
        • What It Is
          • Bridging the lack of modelling capacity
          • Calculator, Code Executor
      • Writing Action
        • What It Is
          • Modify the external environment
          • Modify files, send mail, update data Library
    • What It Is
      • External capability interface for Agent
    • How to Call
      • Function Calling
        • Process
          • Modelled Tool Call
          • Run-on-dol
          • Return Tool Result
          • Add the result to Context
  • Planning
    • What It Is
      • Selection of actions based on objective, constraints and state of the environment
    • Planning strategy
      • Separation of planning and implementation
      • Planned particle size
      • Layer planning
    • Rethinking and miscorrection
      • Inspection plan
      • Check implementation results
      • Reprogramming as necessary
• Harness

How to Optimize It

• Common Failure Modes
  • Planning Failures
    • Error selection tool, parameter error, target not completed
  • Tool-Use Failures
    • Error in tool execution, return error results, lack of necessary tools
  • Efficiency and Cost Problems
    • Too many steps, too long implementation time, too much cost High