Specifying the Agent's Goal

Once you've given your agent its initial set of instructions or its persona, as discussed in "Instructing Your Agent for a Task," the next step is to give it a specific job to do. This job is the agent's goal. Think of the initial instructions as setting the stage, and the goal as defining the main action that needs to happen on that stage. For your first agent, keeping this goal clear and simple is fundamental for success.

Why a Clear Goal Matters

A well-defined goal is the compass for your agent. Without it, your agent, even with the best LLM at its core, might not perform usefully or predictably. Here’s why specifying a clear goal is so important:

1. Direction for Action: The goal tells the agent precisely what outcome you expect. If you want your agent to add an item to a list, a goal like "Update the list" is too vague. A goal like "Add 'Buy bread' to the to-do list" provides clear direction.
2. Focus: A specific goal helps the LLM focus its capabilities on the task at hand, rather than getting sidetracked or providing overly general responses.
3. Measurability: With a clear goal, you can more easily determine if the agent has succeeded or failed. Did it add 'Buy bread' to the list, or did it do something else?
4. Simplicity in Early Stages: For your first agent, simple, unambiguous goals make it easier to build, test, and debug. You'll know exactly what you're aiming for.

Goal vs. Initial Instructions

It's useful to distinguish between the agent's overall instructions (often part of a system prompt or initial setup) and the specific goal for a given task.

• Initial Instructions often define the agent's role, personality, constraints, or general mode of operation. For example: "You are a helpful assistant that manages a to-do list. Respond only with the updated list or a confirmation of the action."
• The Goal is the specific task you want the agent to accomplish right now using those instructions as context. For example: "Add 'Finish project report' to the to-do list."

The instructions provide the "how" and "who" for the agent, while the goal provides the "what."

Characteristics of a Good First Agent Goal

When you're starting out, aim for goals that are:

• Specific and Unambiguous: The goal should state precisely what needs to be done. Avoid vague language. Instead of "Handle my tasks," a better goal would be "Read all tasks due today."
• Achievable: The goal should be something your basic agent, with its current LLM and (perhaps very limited) tools, can realistically accomplish. Don't ask your first simple agent to "Plan a multi-city international vacation." A more achievable goal might be "Suggest three activities for a day trip to San Francisco."
• Atomic: For your first agent, focus on single, discrete tasks. If a task involves multiple steps, break it down and give the agent one goal at a time. For instance, instead of "Organize my to-do list by priority and then email it to me," start with "List all items on the to-do list."

Formulating Effective Goals

Defining a goal for your agent is much like giving a very clear, direct order. Here are a few tips:

1. Start with an action verb: "Add," "Create," "List," "Summarize," "Find," "Remove." This immediately clarifies the desired action.
2. Be explicit about objects and parameters: If the agent needs to act on something or use specific information, include it.
   • Instead of: "Add a task."
   • Use: "Add the task 'Water the plants' to the 'Home' to-do list."
3. Consider the desired output or state change: What should be different after the agent completes its goal? Knowing this helps you phrase the goal effectively.

Examples: Goals for a To-Do List Agent

Later in this chapter, you'll build a to-do list management agent. Let's think about some specific, achievable, and atomic goals for such an agent:

• Adding an item: "Add 'Submit expense report by 5 PM' to the to-do list."

  • This is specific: It states the exact text of the to-do item.
  • It's achievable for a basic agent that can manipulate a list.
  • It's atomic: It's a single operation.

• Viewing items: "Display all items currently on the to-do list."

  • Specific: Clearly asks for a display of all items.
  • Achievable: The agent needs to access and present the list.
  • Atomic: One action to retrieve and show.

• Removing an item: "Remove 'Call the plumber' from the to-do list."

  • Specific: Identifies the exact item for removal.
  • Achievable: Requires finding and deleting an item.
  • Atomic: A single deletion operation.

• Marking an item as complete (a slightly more complex state change, but still manageable): "Mark the task 'Buy groceries' as completed."

  • Specific: Targets a particular task and a state change.
  • Achievable: Involves finding an item and updating its status.
  • Atomic: One update operation on a specific item.

Notice how these goals are direct commands. They leave little room for misinterpretation, which is exactly what you want when guiding your first LLM agent. As you define the goal for your agent, always ask yourself: "Is this crystal clear? Is it a single, manageable step?" If the answer is yes, you're on the right track to building an agent that does what you intend. This clarity will make the subsequent steps of coding and testing your agent much more straightforward.