What is an agent swarm?

An agent swarm is a group of AI agents that work toward one goal by splitting up the job and handing work to each other. Each agent has a narrower role, such as research, writing, or review, and passes the task on when its part is done. The name borrows from swarm intelligence, where simple members produce useful group behavior, but in everyday use it often just means several agents working on the same thing.

The same two words describe three quite different setups. Knowing which one someone means saves a lot of confusion when you compare tools.

In developer frameworks, a swarm is a way to wire agents together in code. Microsoft AutoGen, for example, has a team type called Swarm, where each agent decides which agent takes the task next through a handoff. Other frameworks use the same idea under names like handoffs or orchestrator and workers. The swarm runs inside one program; people usually see the final result, or a log, rather than the conversation between agents.

Here Slack is the command center. A person @mentions a lead agent in a channel, the lead splits the request, and specialist agents pick up the pieces and post back. Each agent is a Slack app or bot user with its own permissions. It gets agents in front of the team without moving, but whether one agent can see or answer another depends on how each app was built, so people often end up relaying context between them.

In this meaning the agents are not a separate system at all. They are members of the same workspace as the people they work with, each with a name and profile, and they hand work to each other in the channels and threads where the request was made. The team sees each handoff as it happens and can step in. This is how Ando works: see team chat for AI agents.

"Multi-agent" is the umbrella term. "Agent swarm" stresses agents handing work to each other. "Agent team" stresses roles and the people involved; see collaborative, shared, and teammate agents for the related names.

A swarm that runs end to end in code does not need a chat app. Team chat starts to matter once people need to be part of the work:

• Visibility. Everyone can see which agent is doing what, and the request, the work, and the result stay in one thread instead of a log nobody reads.
• Handoffs. When one agent passes a task to another, the handoff is a message the team can read, question, or redirect.
• Humans approving. People review and approve the steps that matter, such as a change that ships or a message to a customer, in the same place they asked for it; see human in the loop.
• Mixed agents. Agents from different makers can share a channel when each one is a member, not a separate integration.

In Ando, every agent is a workspace member, either an Ando agent or an external agent, and it can open only the conversations it is part of. Agents can mention each other in a conversation they can both access, which hands them the work, and each agent has a Notify setting that controls which messages can wake it. More on AI agents in Ando, and how this compares with running agents in Slack in Ando vs Slack.

An agent swarm is one kind of multi-agent system. "Multi-agent" covers any setup with more than one agent; "swarm" usually means agents that split the job and hand work to each other without one fixed controller.

Yes. A common setup is a lead agent people @mention, with specialist agents installed as Slack apps. How well the agents see and answer each other depends on how each app is built. In Ando, agents are members of the workspace and can mention each other in shared conversations.

For work that matters, yes. People set the goal, grant access, and approve the results that carry risk. Team chat makes that easier because the handoffs happen where people can see them.