execution_plan

Execution Plan Example

This example demonstrates how to use the execution plan functionality in the Agent SDK. The execution plan feature allows an agent to present a plan of actions to the user before executing them, giving the user the opportunity to approve, modify, or cancel the plan.

Overview

The execution plan functionality provides several benefits:

1. Transparency: Users can see what actions the agent intends to take before they are executed.
2. Control: Users can approve, modify, or cancel the plan before execution.
3. Safety: Prevents the agent from taking unexpected or unwanted actions.

How It Works

1. The agent generates an execution plan based on the user's request.
2. The plan is presented to the user for review.
3. The user can:
   • Approve the plan (by typing "approve" or "yes")
   • Modify the plan (by typing "modify" or "change" followed by the desired modifications)
   • Cancel the plan (by typing "cancel" or "no")
4. If approved, the agent executes the plan and reports the results.

Implementation Details

The execution plan functionality is implemented in the following packages:

• pkg/executionplan: Contains the core execution plan functionality as an independent package.
• pkg/agent: Uses the execution plan package for planning and execution functionality.

The key components are:

• executionplan.ExecutionPlan: A struct that represents a plan of tool executions.
• executionplan.ExecutionStep: A struct that represents a single step in an execution plan.
• executionplan.Generator: Handles generating and modifying execution plans.
• executionplan.Executor: Handles executing plans.
• executionplan.Store: Manages storage and retrieval of plans.

The Agent provides methods to interact with execution plans:

• agent.GenerateExecutionPlan: Generates an execution plan based on user input.
• agent.ModifyExecutionPlan: Modifies an execution plan based on user feedback.
• agent.ApproveExecutionPlan: Approves and executes a plan.

Running the Example

To run this example:

go run cmd/examples/execution_plan/main.go

The example uses a mock LLM for demonstration purposes. In a real application, you would use a real LLM provider like OpenAI.

Example Conversation

Here's an example of how a conversation might flow:

User: What's the weather in New York and convert the temperature to Celsius?

Agent: I've created an execution plan for your request:

# Execution Plan: Search for weather information and calculate temperature conversion

## Step 1: Search for current weather information
Tool: search
Input: current weather in New York
Parameters:
- param1: weather

## Step 2: Convert temperature from Fahrenheit to Celsius
Tool: calculator
Input: convert 75F to C
Parameters:

Would you like to approve, modify, or cancel this plan?

User: modify I want to use the weather tool directly instead of search

Agent: I've updated the execution plan based on your feedback:

# Execution Plan: Search for weather information and calculate temperature conversion

## Step 1: Get current weather information directly
Tool: weather
Input: New York
Parameters:
- param1: New York

## Step 2: Convert temperature from Fahrenheit to Celsius
Tool: calculator
Input: convert 75F to C
Parameters:

Do you approve this plan? You can modify it further if needed.

User: approve

Agent: Executing the approved plan...

Step 1: Get current weather information directly
Result: Executed weather with input: New York

Step 2: Convert temperature from Fahrenheit to Celsius
Result: Executed calculator with input: convert 75F to C

Customization

You can customize the execution plan functionality by:

1. Setting requirePlanApproval to false to disable execution plan approval.
2. Creating your own implementation of the executionplan.Generator or executionplan.Executor.
3. Extending the executionplan.ExecutionPlan and executionplan.ExecutionStep structs with additional fields.

Integration with Your Application

To integrate execution plan functionality into your application:

1. Create an agent with WithRequirePlanApproval(true).
2. Handle user responses to execution plans in your application's UI.
3. Call the appropriate methods (ApproveExecutionPlan, ModifyExecutionPlan) based on user responses.
4. Directly use the executionplan package if you need more control over the planning and execution process.