Grapgflow is not a parallel flow, actually just a concurrent flow

[prety-good]

What happened?

Describe the bug
I am glad to see the launch of graphflow function，but after my attempts, Grapgflow is not a parallel flow, it is actually just concurrent. It just executes the agents that need to be parallel sequentially.

To Reproduce
In the graphflow function, I create 3 test_agent to parallelly run:

# Build graph with conditional loop
builder = DiGraphBuilder()
builder.add_node(begin_agent).add_node(test_agent1).add_node(test_agent2).add_node(test_agent3).add_node(over_agent)
# 1
builder.add_edge(begin_agent, test_agent1)
builder.add_edge(begin_agent, test_agent2)
builder.add_edge(begin_agent, test_agent3)
# 2
builder.add_edge(test_agent1, over_agent)
builder.add_edge(test_agent2, over_agent)
builder.add_edge(test_agent3, over_agent)

builder.set_entry_point(begin_agent) 
self.flow = GraphFlow(
    participants=builder.get_participants(),
    graph=builder.build(),
)

While in my code, I use the asyncio.gather to parallelly run, just like:

test_agent1 = TestAgent(self.model_client, name="test_agent_1")
task1 = RoundRobinGroupChat(
    participants=[test_agent1],
    termination_condition=(
        TextMentionTermination("None") |
        MaxMessageTermination(max_messages=30) |
        TextMentionTermination("已完成") |
        TextMentionTermination("TERMINATE")
    )
)
test_agent2 = TestAgent(self.model_client, name="test_agent_2")
task2 = RoundRobinGroupChat(
    participants=[test_agent2],
    termination_condition=(
        TextMentionTermination("None") |
        MaxMessageTermination(max_messages=30) |
        TextMentionTermination("已完成") |
        TextMentionTermination("TERMINATE")
    )
)
test_agent3 = TestAgent(self.model_client, name="test_agent_3")
task3 = RoundRobinGroupChat(
    participants=[test_agent3],
    termination_condition=(
        TextMentionTermination("None") |
        MaxMessageTermination(max_messages=30) |
        TextMentionTermination("已完成") |
        TextMentionTermination("TERMINATE")
    )
)
swarms = [task1, task2, task3]
consoles = [Console(swarm.run_stream(task=messages)) for swarm in swarms]
await asyncio.gather(*consoles)

Expected behavior
In the parallelism implemented by graphflow, the outputs of different agents are independent and appear to be executed sequentially.
In my code, the outputs of different agents are mixed and the running speed is twice as fast as graphflow.
Is there something wrong with my implementation, or is graphflow just a concurrent method and does not support parallelism?

Which packages was the bug in?

Python AgentChat (autogen-agentchat>=0.4.0)

AutoGen library version.

Python 0.5.6

Other library version.

No response

Model used

No response

Model provider

None

Other model provider

No response

Python version

3.10

.NET version

None

Operating system

MacOS

[ekzhu]

@prety-good thanks for your issue. It is correct the GraphFlow currently executes nodes sequentially. The API is experimental so consider it not fully baked. There are several things under development including support for functional conditions in edges.

If you are interested to discuss implementation, you can join Discord channel: https://discord.com/channels/1313229474107756584/1341892471084548198 for contributors.

Converting this to a discussion.

[prety-good]

Thanks for your reply. Actually, I’m very interested in discussing implementation — for example, adding support in GraphFlow for agents to run truly in parallel. Compared to concurrency, parallelism may lead to disordered output formatting, but it can significantly improve GraphFlow’s execution speed — which is critically important in my own project.
So, how should I get started if I’d like to contribute?