SEC MCP

# Agents vs. Workflows in AI This document summarizes key points from the transcript regarding AI agents and workflows, along with illustrative diagrams. --- ## 1. What Are Agents? An **agent** in the AI context is an autonomous system where the LLM (Large Language Model): - **Decides its own path:** It determines how many iterations or steps are needed. - **Operates dynamically:** It can loop through actions until a satisfactory resolution is achieved. - **Utilizes multiple tools:** It might perform tasks such as web searches, code execution, or file editing as part of its decision process. *Example use cases include customer support interactions and iterative code improvements.* --- ## 2. Workflows vs. Agents ### Workflows A **workflow** is a pre-defined, fixed sequence of steps where: - **Each step is predetermined:** The process flows from one prompt to the next. - **The number of steps is known upfront:** It’s like a rail that guides the data transformation. - **The execution is linear:** For example, categorizing a user query before processing it further. ### Agents An **agent**, in contrast, is characterized by: - **Autonomy in decision-making:** The LLM chooses the next action. - **An iterative loop:** The process continues until the goal is reached. - **Flexibility:** It may decide to run different actions (like performing a web search or editing code) based on the current context. --- ## 3. Workflow Prompt vs. Agent Prompt ### Workflow Prompt A workflow prompt follows a fixed, linear progression: ```mermaid flowchart TD A[User Input] --> B[Prompt A] B --> C[Prompt B] C --> D[Prompt C] D --> E[Final Output] ``` ### Key Characteristics: Fixed sequence: The output of one prompt feeds directly into the next. Predictable flow: Designed with a known number of steps. ## Agent Prompt ### An agent prompt is open-ended and iterative, allowing the system to choose among various actions until it reaches a resolution: ```mermaid flowchart TD A[User Query] --> B[Agent] B --> C{Resolution Reached?} C -- Yes --> D[Final Answer] C -- No --> E[Choose Action] E --> F[Web Search] E --> G[Run Code] E --> H[Edit File] F --> B G --> B H --> B ``` #### Key Characteristics: Dynamic decision loop: The agent continues taking actions until a satisfactory answer is found. Tool selection: The agent can pick from multiple actions based on context. # Empathy and Clarity: How To Design Agents This section highlights personal experiences and insights into designing effective agents by understanding the model’s perspective and the importance of clear tool descriptions. --- ## 1. Learning Through Experience One engineer shared a personal story about their early experience with agent design during an onboarding task (running a computer use benchmark called OSWorld). For an entire week, they observed agent trajectories that seemed counterintuitive, until they decided to "act like Claude" (the model) by simulating its behavior: - **Simulation Technique:** - Close your eyes for a minute. - Blink at the screen for a second. - Reopen your eyes and think: **"I have to write Python code to operate in this environment, what would I do?"** This method helped them better understand the decision-making process of the agent. ```mermaid flowchart LR subgraph "Agent Simulation Process" A([Observe Agent Behavior]):::action B([Identify Unusual Patterns]):::action C([Act like Claude: Close Eyes & Blink]):::action D([Simulate Python Decision-Making]):::action E([Gain Insight into Agent Logic]):::result end A --> B B --> C C --> D D --> E classDef action fill:#f9f,stroke:#333,stroke-width:2px,stroke-dasharray: 5 5; classDef result fill:#bbf,stroke:#333,stroke-width:2px; ``` ## 2. The Importance of Empathy in Prompt Engineering **A key takeaway is the necessity for developers to step into the model's shoes:** ### Empathetic Perspective: Consider how the model "sees" the world. Its understanding is based solely on the context provided in the prompt. ### **Tool Descriptions Matter:** - Many detailed prompts fail because the tools provided to the model are under-documented (e.g., parameters named only "A" and "B"). - Just like function documentation for engineers, tool descriptions need to be clear and thorough. - Unified Context: - Remember that all parts of the prompt (the main instructions, tool descriptions, and additional context) are fed into the same context window. Each element influences the others. ```mermaid flowchart TD subgraph "Effective Prompt Engineering" A([Craft Detailed Prompt]):::action B([Design Tool Descriptions]):::action C([Provide Clear Documentation]):::action end A --> D([Integrate into Unified Context]):::context B --> D C --> D D --> E([Facilitate Effective Model Interpretation]):::result classDef action fill:#cfc,stroke:#333,stroke-width:2px,stroke-dasharray: 4 4; classDef context fill:#fcf,stroke:#333,stroke-width:2px; classDef result fill:#cff,stroke:#333,stroke-width:2px; ``` ## 3. Developer Tips ### 1. Act Like the Model: **Approach the problem from the model’s perspective. Ask yourself, "How does the model see the world with the given context?"** ### 2. Document Your Tools Thoroughly: **Just as engineers rely on detailed API documentation, your tool descriptions should be comprehensive to ensure the model can use them effectively.** ### 3. Consider the Unified Prompt: **Remember, all parts of your prompt—detailed instructions, tool descriptions, and contextual hints—contribute to the final behavior of the agent.** These insights and diagrams emphasize that while agents are the "hot term" in AI right now, the success of agent-based systems hinges on thoughtful, empathetic design and clear communication within the prompt. ```mermaid flowchart TD A[Developer Tips]:::start A --> B[Act Like the Model] A --> C[Document Your Tools Thoroughly] A --> D[Consider the Unified Prompt] B --> E[Gain Empathetic Perspective] C --> F[Provide Clear, Detailed Documentation] D --> G[Integrate All Prompt Elements] E --> H[Guide Model's Decision-Making] F --> H G --> H H --> I[Effective Agent Behavior] classDef start fill:#f9f,stroke:#333,stroke-width:2px; classDef tip fill:#cfc,stroke:#333,stroke-width:2px,stroke-dasharray: 5 5; classDef result fill:#cff,stroke:#333,stroke-width:2px; ```