Want to StateManagement in GroupChat Orchestration

[Saravanan123ghnsms]

I am working in python,
I want to know whether it is possible to apply state management in group chat orchestration in agent framework.
I have tried in documentation . I didn't found any relevant documentation.

Note : In custom flows it is possbile by Shared State but I want to work on Inbuilt groupchat builder.

Thank you

[moonbox3]

Hi @Saravanan123ghnsms, can you help us understand how you plan to use shared state as part of the group chat builder/pattern?

[Saravanan123ghnsms]

Hi @moonbox3 ,

Thank you for your response.

Let me clearly explain my scenario and how I am planning to use shared state with the Group Chat orchestration pattern.

I have two agents:

Agent 1

Has a single tool

The tool returns a result, which may be either an object or a string

Agent 2

Has a single tool

The tool also returns a result, which may be either an object or a string

Both agents are attached to a Group Chat orchestration.

My requirement

When Agent 1 executes its tool, the tool returns a structured result (for example, an object).
I want Agent 2 to explicitly access Agent 1’s tool result inside its own code, by reading from a shared state or conversation state.

The reason I want to do this programmatically is because relying purely on the LLM to interpret prior messages is not always reliable. The LLM may:

Miss fields

Change formats

Hallucinate values

Instead, I want Agent 2 to directly read Agent 1’s tool output from shared state, so it can safely use that data when invoking its own tool.

Question

Is it possible, when using Group Chat orchestration, to:

Persist Agent 1’s tool result into a shared state

Access that shared state directly inside Agent 2’s code (not via LLM reasoning)

If this is supported, could you please explain the recommended approach?

Thank you.

[Saravanan123ghnsms]

Hi @moonbox3 ,

Thank you for your response.

Let me clearly explain my scenario and how I am planning to use shared state with the Group Chat orchestration pattern.

I have two agents:

Agent 1

Has a single tool

The tool returns a result, which may be either an object or a string

Agent 2

Has a single tool

The tool also returns a result, which may be either an object or a string

Both agents are attached to a Group Chat orchestration.

My requirement

When Agent 1 executes its tool, the tool returns a structured result (for example, an object).
I want Agent 2 to explicitly access Agent 1’s tool result inside its own code, by reading from a shared state or conversation state.

The reason I want to do this programmatically is because relying purely on the LLM to interpret prior messages is not always reliable. The LLM may:

Miss fields

Change formats

Hallucinate values

Instead, I want Agent 2 to directly read Agent 1’s tool output from shared state, so it can safely use that data when invoking its own tool.

Question

Is it possible, when using Group Chat orchestration, to:

Persist Agent 1’s tool result into a shared state

Access that shared state directly inside Agent 2’s code (not via LLM reasoning)

If this is supported, could you please explain the recommended approach?

Thank you.

[moonbox3]

Since we don't have any way, right now, to make the state public from the high-level Group Chat orchestration pattern, I'd recommend looking into rolling your own:

agent-framework/python/packages/orchestrations/agent_framework_orchestrations/_group_chat.py

Lines 1003 to 1020 in c609b14

	participants: list[Executor] = self._resolve_participants()
	orchestrator: Executor = self._resolve_orchestrator(participants)
	# Build workflow graph
	workflow_builder = WorkflowBuilder().set_start_executor(orchestrator)
	for participant in participants:
	# Orchestrator and participant bi-directional edges
	workflow_builder = workflow_builder.add_edge(orchestrator, participant)
	workflow_builder = workflow_builder.add_edge(participant, orchestrator)
	if not self._intermediate_outputs:
	# Constrain output to orchestrator only
	workflow_builder = workflow_builder.with_output_from([orchestrator])
	if self._checkpoint_storage is not None:
	workflow_builder = workflow_builder.with_checkpointing(self._checkpoint_storage)
	return workflow_builder.build()

This is the code that configures the Agents (wrapped as AgentExecutors). In that way, you'd have access to the Executor's ctx: WorkflowContext where you'd then be able to set/get state.

[Saravanan123ghnsms]

Thank You so much @moonbox3 ,

I will do that.