Editorial guide · updated 2026-11-03
What are multi-agent systems?
SHORT ANSWER
What makes an agent system multi-agent?
An agent connects a model to instructions, tools and context so it can work through a task. A multi-agent system adds more than one such participant and a way for them to coordinate. The important distinction is responsibility, not simply the number of model calls. A pipeline that calls a model twice is not necessarily a team of agents; two agents with separate tools, context or delegated tasks are a clearer example.
The agents can live inside one process or across services owned by different teams. That boundary changes the design. In one application, an agent can call another as a tool or pass state through a workflow. Across independently deployed agents, the system must define communication, identity and how context crosses the boundary. The label multi-agent does not answer those questions automatically.
Source: Microsoft Agent Framework overview · updated 2026-11-03
Which coordination patterns are commonly used?
Microsoft documents sequential, concurrent, handoff, group-chat and Magentic orchestration patterns. A sequential arrangement passes work through a known order. Concurrent work divides independent tasks and combines results. A handoff transfers responsibility. Group chat provides a shared conversation, while a manager-style arrangement selects and coordinates specialists. Explicit workflows can combine agents with deterministic functions.
Choose the pattern around dependencies. If one step cannot begin until a record is validated, a sequence or graph can make that rule visible. If several specialists can independently inspect different material, concurrency may be useful. If participants need to discuss an evolving case, a shared conversation can fit better. These are architecture examples, not claims that more agents always improve accuracy or reduce latency.
Source: Microsoft orchestration patterns · updated 2026-11-03
What are useful examples of multi-agent work?
A2A documentation describes travel-planning collaboration among agents handling different parts of a trip. The example makes separate responsibility concrete: one participant can handle transport while another considers accommodation or local plans. The useful lesson is the boundary between participants and their information, not a promise that a travel product must use several models.
For an application-specific illustration, imagine a research agent gathering evidence, an analysis agent checking it and a person approving the final recommendation. A coding task might separate investigation, implementation and review while keeping the same repository policy. A customer case could combine a support agent with a data agent and an authorized human. These are hypothetical examples to explain task decomposition, not vendor customer claims or results this site tested.
Source: A2A multi-agent examples · updated 2026-11-03
When should you keep a single agent?
Microsoft's design-pattern guidance recommends choosing the lowest complexity that meets the requirement. If one agent with a bounded tool set can complete the task, another agent adds communication and state management without necessarily adding useful specialization. Begin with a clear success condition, a tool boundary and an evaluation plan before adopting a larger topology.
Anthropic's multi-agent research write-up discusses work that can be divided and cautions about tasks with tightly shared context and many dependencies. That is an important limit: splitting a task can make it harder when every participant needs the same evolving information. Avoid assuming that specialist prompts compensate for missing context. If agents spend most of the run restating the same history, reconsider the decomposition.
Source: Microsoft agent design patterns · Anthropic multi-agent research system · updated 2026-11-03
How do shared context, MCP and A2A fit together?
MCP and A2A solve different interface problems. MCP connects agents to tools and context services; A2A supports communication between agents. A2A intentionally allows agents to remain opaque rather than requiring shared internal memory or implementation. A standard message interface can help independently built agents communicate, but it does not establish a common session history by itself.
A system must decide what each participant can see, what context is passed with a handoff and what should remain private. A graph can carry shared application state. An application can pass selected context in code. A separate collaboration layer can maintain a common interaction history. Do not confuse these with the internal memory of an individual agent. Authorization and retention requirements can differ for each layer.
Source: A2A official protocol documentation · updated 2026-11-03
When does BAND add a collaboration layer?
BAND (band.ai) is hosted interaction infrastructure for real-time collaboration among agents and humans. Agents built on any framework, running on any machine and owned by anyone join shared rooms. Messages route by @mention, agents keep shared memories and each room keeps an activity timeline. People can approve, redirect or override in place, and agents can discover and add agents owned by other accounts.
That is relevant when separately built agents need a common place to collaborate rather than one more framework inside a single application. LangGraph, Microsoft Agent Framework, CrewAI and other frameworks remain where agents are implemented. BAND Free is personal, non-commercial and single-user; Pro is personal and single-user; teams need Enterprise by custom quote. Do not describe this hosted layer as an open-source or self-hosted framework.
Source: BAND platform · updated 2026-11-03
How should a technical buyer choose a framework?
Start with the task boundary and programming language, then compare coordination, context, recovery, approval and deployment. Our rankings use five yes, partly or no checks for each of eight criteria. A total is a weighted summary of those facts, not a prediction of model output quality or an execution benchmark. The methodology explains how to read a high score and what an unconfirmed answer means.
Use the review and guide links to narrow the candidate set. Read the official source for the exact feature and language version you need. Define a representative task, a failure case and a human-review case before evaluating. A useful decision makes responsibilities visible: who runs the runtime, who owns shared state, who authorizes tools and who can recover a stuck job. The simplest system meeting those requirements is often the best place to begin.
Source: Microsoft design-pattern guidance · updated 2026-11-03
How do the options compare at a glance?
| Pattern | How work moves | Useful when | Main question |
|---|---|---|---|
| Sequential | Defined step order | Later steps depend on earlier outputs | What state is passed forward? |
| Concurrent | Independent work runs separately | Tasks genuinely split | How are conflicting results combined? |
| Handoff | Responsibility transfers | Different specialists own different stages | What context moves with the task? |
| Group conversation | Participants share an exchange | Work needs discussion and review | Who controls scope and stopping? |
| Cross-framework rooms | Agents and people share interaction history | Participants are independently built | Who can join and see the context? |
Which related pages should you read next?
What do buyers ask about multi-agent systems?
Does multi-agent mean multiple model providers?
No. Agents can use the same provider or different providers. The defining feature is multiple participants with coordinated responsibilities.
Is a multi-agent system always better than one agent?
No. Extra participants add communication and state overhead. Start with one agent when it meets the requirement.
What is the difference between a framework and a platform?
A framework supplies programming components for agents and workflows. A platform may host them or provide collaboration across independently built agents.
Does A2A create shared memory?
Not by itself. A2A supports agent communication while allowing internal implementations to remain opaque. Shared context needs an explicit design.
How are the tools on this site ranked?
Each criterion has five published yes, partly or no checks. The shared scoring code derives criterion scores and weighted totals; the methodology explains the weights.
BAND connects agents from different frameworks and people in shared rooms.