Skip to main content
    Kriv AI

    Agentic AI design

    Multi agent workflow: the five patterns and when you need one

    A plain guide for owners and IT leads, based on Microsoft Learn and Anthropic, read 11 November 2026.

    A multi agent workflow is a process in which several AI agents share the work, each with its own role, instead of one agent doing everything. Microsoft's Azure Architecture Center, read on 11 November 2026, describes five patterns: sequential, concurrent, group chat, handoff and magentic. It also advises using the lowest level of complexity that reliably meets your requirements, because multiple agents add coordination overhead, latency and failure modes. Start by asking whether a single agent with tools can do the job. If it cannot, pick the pattern that matches how your work actually flows, and decide in advance where a person approves.

    plain answer

    What is a multi agent workflow?

    A multi agent workflow is a process where more than one AI agent takes part, each handling a part of the task. Microsoft's guide says that splitting a complex problem across dedicated agents with specific capabilities mirrors strategies found in human teamwork, and lists specialization, scalability, maintainability and optimization as the advantages over a single large agent.

    Anthropic's article "Building effective agents", read on the same day, separates two ideas. Workflows are systems where models and tools are orchestrated through predefined code paths. Agents are systems where models dynamically direct their own processes and tool use. A multi agent workflow can use either idea, which is why the pattern you choose matters.

    how to

    When is one agent enough for a multi agent workflow?

    For most teams the first question about a multi agent workflow is whether you need one. Microsoft sets out three levels. A direct model call suits single-step tasks such as classification or summarization. A single agent with tools suits varied requests in one domain, and Microsoft calls it often the right default for enterprise use cases, simpler to debug and test than multiagent setups. Multiagent orchestration suits cross-domain problems, tasks that need distinct security boundaries for each agent, and work that benefits from parallel specialization.

    Microsoft says plainly that each level adds coordination overhead, latency and cost, and that multiagent setups add failure modes. Anthropic gives the same advice: find the simplest solution possible and increase complexity only when needed, which might mean not building an agentic system at all.

    how to

    What are the five multi agent workflow patterns?

    Microsoft's guide names five orchestration patterns for a multi agent workflow. Each description below follows its page.

    Sequential: agents are chained in a predefined, linear order, and each processes the output of the one before. The order is set in the workflow, not chosen by the agents. It suits steps with clear dependencies, such as draft, review, polish.

    Concurrent: multiple agents run at the same time on the same task, each giving an independent view. The page also calls this fan-out and fan-in.

    Group chat: agents work in a shared conversation, and a chat manager decides which agent responds next. It can support brainstorming or structured quality gates.

    Handoff: an agent decides whether to handle a task or pass it to a better-suited agent, and full control moves to the next agent. It suits cases where the right specialist is not known upfront.

    Magentic: for open-ended problems with no fixed plan. A manager agent builds a task ledger of goals and subgoals and refines it as the agents work. Microsoft notes its agents typically have tools that make direct changes in external systems.

    how to

    How do you choose a multi agent workflow pattern?

    Pick the pattern that matches how the work already flows. If the steps are known and in order, sequential is the simplest fit. If independent views on the same input are useful, concurrent fits. If the right expert only becomes clear during the work, handoff fits. Microsoft says to avoid handoff when the right agent is identifiable from the initial input, and to use a simpler dispatcher that classifies the input and sends it on.

    In our reading, Anthropic's orchestrator-workers pattern, in which a central model breaks a task down, delegates to worker models and combines the results, is the closest match to Microsoft's magentic idea. Our suggestion is to use the more open patterns last, since they give agents more freedom.

    how to

    What should you control in a multi agent workflow?

    Our suggestion, which is Kriv's guidance and not a statement from either source: write down, for each agent in the workflow, what it may read, what it may change, and who approves a change that matters. Microsoft's own list of reasons to use several agents includes distinct security boundaries for each agent, which points to the same habit.

    Microsoft also advises setting iteration limits to guard against infinite tool-call loops. Keep a record of which agent did what, so a person can review a result afterward, and put a human approval step before any action that sends, pays, changes a record or acts outside the company.

    kriv role

    Where Kriv helps

    Kriv AI helps teams scope, set up and govern agentic AI work, including where a person approves. See agentic AI architecture, orchestration platform governance and agentic AI training for how that work is described on this site. Kriv did not write either source and neither company has reviewed this page.

    Your own technical and risk leads hold the final say.

    limits

    What this page could not confirm

    We read Microsoft's pattern guide (page dated 12 February 2026 in its header, with an update stamp of 21 September 2026) and Anthropic's article once each on 11 November 2026. We did not test any agent framework and make no claim that a pattern gives better results for your work. Pattern names differ between vendors, and products change, so check the current pages.

    Straight answers

    Frequently asked questions about Multi agent workflow: the five patterns and when you need one

    What is a multi agent workflow?

    A process in which several AI agents share a task, each with a defined role. Microsoft's Azure Architecture Center, read 11 November 2026, documents five ways to organize them.

    Do we need several agents?

    Often not. Microsoft calls a single agent with tools often the right default for enterprise use cases, and says to use the lowest level of complexity that reliably meets your requirements.

    What is the difference between a workflow and an agent?

    Anthropic defines workflows as systems where models and tools are orchestrated through predefined code paths, and agents as systems where models dynamically direct their own processes and tool use.

    Which pattern should we start with?

    The one that matches how your work flows. Sequential is the simplest when the steps are known and in order. The more open patterns, such as group chat and magentic, give agents more freedom and need more care.

    Can Kriv help us design one?

    Kriv AI helps teams scope, set up and govern agentic AI work. Your own technical and risk leads decide what is built.

    Talk to the team that would do the work

    Bring your requirements to a working session with the person who'll actually deliver.

    Book a Discovery Call