How to store history for a workflow?

[jooooel]

Hi,
I have a ChatClientAgent in an API endpoint. A request comes in with a question, an agent is created and returns the answer to the question.

In order to keep history and allow follow up questions, I'm serializing and deserializing the ChatClientAgentThread for each request so that the agent has the full history. This works fine.

I'm now trying to move to a kind of orchestrator pattorn, using a workflow with handoff instead. This gives me a WorkflowThread. The problem with that thread is that it can't be serialized (https://github.com/microsoft/agent-framework/blob/9148392d00dafa47a95178622f3800f89bbf0425/dotnet/src/Microsoft.Agents.AI.Workflows/WorkflowThread.cs).

Is this by design or just not implemented yet? Are there other more recommended approaches to solve this type of issue (restoring history/context between requests).

It also seems like the WorkflowThread only contains the original question. Not the reply and not any messages of the sub agents. Is something else keeping the full history, or is the idea that each agent should store and restore their own thread history? What's the best approach to pass some kind of thread id to each agent in that case?

This is .NET using 1.0.0-preview.251009.1.

[moonbox3]

Adding @stephentoub for help.

[stephentoub]

@westey-m / @lokitoth ?

[jooooel]

No? 😅 😁

[westey-m]

@jooooel, I'll defer to @lokitoth on the WorkflowThread specifics, but just calling out that this is work in progress. See the recent update #1498

More generally, the intended design is to allow any AIAgent and AgentThread (even ones that wrap a more complex process) to be used in the same way, including allowing serialization and persistence of the AgentThread. While different AgentThread types from different AIAgent types may store different data, or even store data externally with just a reference in the AgentThread, a consumer would be able to follow the same "read-deserialize-run-serialize-save" process for all.

[Zochory]

Also using the conversation OpenAI format like in DevUI for example can potentially be useful for you !