Regular vs Swarm vs Workflow vs Graph Agents

[rmadabusiml]

I sort of understood what and when to use these type of agents but after looking at the https://github.com/strands-agents/samples/blob/main/02-samples/09-finance-assistant-swarm-agent/finance_assistant_swarm.py, I kind of lost my way of understanding Swarm.

Can someone explain difference between Swarm vs Workflow vs Graph and when to use each?
How are conversation history and memory shared between these type of agents?
Note: I went thru the documentation as well but since it uses same examples for all of these, I'm unable to understand when to use what and the above finance_assistant_swarm.py threw me off totally

[flowconic]

following

[yonib05]

Hi @rmadabusiml, thanks for the question, and sorry it took us this long to get you a proper answer. The good news is the docs now have exactly the comparison you were asking for: https://strandsagents.com/docs/user-guide/concepts/multi-agent/multi-agent-patterns/

The short decision guide: the key question is who decides the path of execution. A single Agent when one agent with the right tools can do the job, always start here. A Graph when you the developer define the structure up front (nodes, edges, conditions, loops) and want controlled, predictable routing, like intent-based support flows or validation with error paths. A Swarm when the agents themselves decide, handing off to whichever specialist peer fits next; the path is emergent, which suits exploration and collaborative tasks. And Workflow when you have a fixed DAG of tasks with dependencies, deterministic, no cycles, independent tasks running in parallel. One nuance: unlike Graph and Swarm, Workflow is a pattern you implement rather than a built-in class.

On history and memory: each agent always keeps its own conversation history. Across agents, a Swarm shares a working context (the original request, handoff history, and knowledge earlier agents contributed) with every agent; a Graph passes each node's output along its edges; a Workflow task only receives the results of the tasks it depends on. Graph and Swarm additionally accept an invocation_state dict that's shared with every agent and tool without ever entering the prompt, which is the right place for things like session IDs or connections.

On the finance assistant sample that threw you off: fair, the version you read was written before the built-in Swarm shipped. It's since been rewritten around strands.multiagent.Swarm and moved to python/04-industry-use-cases/finance/finance-assistant-swarm-agent/ in the samples repo (the old link is dead). The Swarm docs are the best reference either way: https://strandsagents.com/docs/user-guide/concepts/multi-agent/swarm/

@flowconic, the pages above should cover what you were following for too. Hope this clears it up!