[Roadmap] Persistence and state management

[ekzhu]

We want to be able to persist the state of a multi-agent workflow and restart from a given state. This is required for large-scale deployment of AutoGen workflows. The key steps are:

1. How to inject state (e.g., conversation history) into a multi-agent workflow so it starts from a given state. This functionality can also be used to implement in-context learning for individual agents, so they learn from a curated set of instructions and examples before participating in a workflow.
2. How to export the state of a multi-agent workflow.

We start building (1) and (2) for individual class:

• single-agent
• two-agent chat
• sequential chats
• group chat,

and then move on to think about a composed workflow of multiple conversation patterns.

Tasks

Beta Give feedback

1. GPTAssistantAgent add "thread_id" to its constructor to gain prior message histories. @jtrugman
2. initiate_chat to add prior conversation histories for the two-agent chat.
3. [Feature Request]: Resumable GroupChat #2359
   enhancement group chat +2
4. Inject message history through initiate_chat. #2671
   enhancement help wanted persistence +3

[NikolayTV]

Hi! I have following use case:
In multi-agen conversation with human, there is an Agent named "collector" which collects info about the user to json from conversation. That meta-information I want to use sometimes, and "coordinator" will decide when to call it to context.

Is this feature about state management? I want to implement it for my project, and may be it will be nice to PR this to main repo too?

For example:
user_meta_info = {
"Name": "Boris",
"Chronic deseases": "astma, headache"
}

[ekzhu]

You are thinking about a usage case that is a natural language text to JSON conversion. You can certainly create a two-agent chats to collect human input and then in the initiate_chat use a summary_method=reflection_with_llm and summary_args={"summary_prompt": "extract the following information ... and returns a JSON"}. This is not directly really related to the state management.

[randombet]

How could this support async chats?