Task Loop

​

What is the Task Loop?

The task loop is SpinAI’s core decision-making process that orchestrates your AI agent’s workflow. It handles:

• Planning which actions to run and when
• Determining parameters for each action
• Executing actions in parallel when possible
• Generating reasoning and final responses
• Maintaining state throughout the interaction
• Comprehensive logging and cost tracking

​

How it Works

The task loop uses a sophisticated planning system that leverages your agent’s LLM for key decisions:

1. Planning and Reasoning

   • Analyzes user input, current state, and previous action results
   • Provides explicit reasoning about what to do next
   • Decides which actions should run next (if any)
   • Returns a list of action IDs with their parameters

2. Parallel Action Execution

   • Executes multiple actions in parallel when possible
   • Each action receives the full context and specific parameters
   • Results from each action are captured and added to the conversation history
   • State is maintained and accessible across all actions

3. Response Generation

   • When no more actions are needed, generates a final response
   • Uses accumulated context, action results, and state
   • Can return plain text or structured data based on your responseFormat configuration
   • Provides detailed usage metrics (tokens, cost, duration)

​

Flow Diagram

​

Technical Implementation

The task loop is implemented with these key components:

1. Message History Management

   • Maintains a structured conversation history with system, user, assistant, and action result messages
   • Each message type has a specific purpose in the planning process

2. Planning Schema

   • Uses a structured JSON schema for planning decisions
   • Includes reasoning, next actions, and parameters for each action

3. Error Handling

   • Gracefully handles errors in action execution
   • Provides error information back to the agent for recovery

4. Comprehensive Logging

   • Tracks planning decisions, action executions, and final responses
   • Records performance metrics including token usage and costs

​

Next Steps