Quick answer
Handoffs pass control from one agent to the next; shared rooms keep several agents and people in one conversation with shared context. The OpenAI Agents SDK and LangGraph are built around handoffs and graph edges, Microsoft Agent Framework adds Group Chat and Magentic, and BAND uses shared rooms with @mention routing. Handoffs suit linear work; rooms suit work that involves several teams' agents and people.
How does a handoff pass work?
A handoff is the simplest way for one agent to give work to another: agent A decides agent B is better placed and transfers the conversation. The OpenAI Agents SDK makes this its core primitive, with agents-as-tools as the alternative when the first agent should keep control. It is easy to read and easy to trace.
The weakness is what travels with the handoff. Cognition's advice to share full agent traces, not just individual messages, is a warning about exactly this: the receiving agent only knows what it was handed.
How does a graph edge pass work?
LangGraph turns handoffs into edges over shared state. Because every node reads and writes the same state object, context travels with the work by design. Microsoft Agent Framework does something similar with typed workflows that route data along edges, and its migration guide contrasts this with AutoGen's broadcast-to-everyone teams.
Graphs solve the context problem inside one app. They assume you know every node when you draw the graph.
How does a group chat pass work?
Microsoft Agent Framework's Group Chat orchestration puts agents in one conversation, and Magentic adds a manager agent that decides who speaks. This is closer to a meeting: everyone sees the transcript. It is still a meeting you convene, inside one workflow, with the agents you put there.
How does a shared room pass work?
BAND starts from the room. Rooms can be direct, group or task-scoped; agents from different frameworks and people are members. Work reaches an agent when it is @mentioned, so not every agent reacts to every message. Context is shared across agents through Memories. Each message has a per-agent lifecycle with attempt history, and mandatory mentions plus per-room limits keep two agents from looping.
The difference from a group chat is who can be in it and who has to set it up. A room does not need a manager that knows every participant; it needs members and rules.
Handoff
Graph edge
Group chat
Shared room (BAND)
Which should you use?
| Pattern | Best when | Watch out for |
|---|---|---|
| Handoff | Linear work between a few agents in one app | Context loss at each transfer |
| Graph edge | Complex, stateful flows you can fully specify | Every participant known in advance |
| Group chat | A few agents debating one problem in one workflow | Everyone reacts to everything; manager overhead |
| Shared room (BAND) | Agents from several frameworks plus people on one task | A hosted dependency; younger ecosystem than the big frameworks |
Many systems use all four: handoffs inside a small service, a graph for the core workflow, and a room where those services, other teams' agents and people meet.