Template Patterns

Long-Running Agent with Compaction

Describes a persistent conversational-agent pattern that compacts chat history to manage long-term context usage.

This simple pattern involves a conversational agent in a single state that has many actions that it can use to fulfill the users request.
This pattern is typically initiated once (sometimes a setup state is used) and then handles tasks for a user over a long period of time.
This means the conversational chat needs to implement chat squashing to compact the context window for long term usage.

# Examples

• A home shopping list manager
• A task and time logging agent for employees of a consultancy firm
• An agent that takes instructions to add and check off tasks in a project management system

# Detailed Structure

• An agent is initiated on the _create action
• There is an optional 'setup-in-progress' state which sets up the basic details of the agent with input from the user
• An active state has a collection of actions that the agent can use.
• Replies from the AI agent are routed to a notification platform e.g. Whatsapp
• The stubsession is set to timeout in a year ~9000 hours
• This allows the user to keep the chat open to the stub and access the agent at any time
• On each update from the AI Agent check to see if we need to squash the chat, i.e. Compact the context window
• The gpt_chat_task safeguards need to be disabled