Task 5.1: Operate and manage multi-agent workflows

• ARefreshing documentation where each worker owns one page

  Incorrect

  Documentation refreshes with one owner per page are listed as good fits.

• BReviewing a large set of diffs, one module per worker

  Incorrect

  Batch reviews with a separate module per worker are listed as good fits.

• CA migration where step two needs the concrete output of step one

  Correct

  Sequential tasks where a step needs the previous step's output are listed as poor fits.

• DResearching the same question across five independent services

  Incorrect

  Parallel research across independent services is a good fit.

• EA change in which every worker would need to edit the same three core files

  Correct

  Tightly coupled edits where workers contend for the same files are listed as poor fits.

Fleet mode suits work that can be decomposed before execution, with each unit able to run without waiting for others: multi-file refactors with one owner per file or package, batch reviews, parallel research across independent services, documentation refreshes and migrations where each worker validates its own slice. Avoid it for sequential tasks, tightly coupled edits on the same files, small tasks and work that needs continuous shared reasoning.

• AThe /fleet command, which repeats the same split of work every run

  Incorrect

  With /fleet, Copilot decides how to divide and coordinate the work each time.

• BAutopilot mode, which keeps a fixed sequence of steps across runs

  Incorrect

  Autopilot is about autonomy: Copilot decides the next steps itself.

• CA custom agent whose prompt lists the steps in order for the main session to follow

  Incorrect

  A prompt guides one agent's behavior; it doesn't fix the stages and conditions the way code does.

• DA dynamic workflow, whose author defines the steps, conditions and handoffs in code

  Correct

  In a dynamic workflow the process is defined in code by its author, and agents handle the parts that need judgment.

Autopilot lets Copilot keep working without pausing for input, with Copilot deciding the next steps. /fleet encourages Copilot to delegate to subagents in parallel, deciding how to break the task down each time. A dynamic workflow carries out a process defined in code: its author defines the steps, conditions and handoffs, and agents handle the parts that need analysis or judgment.

• AIt works, as long as each subagent uses a different model from the main session

  Incorrect

  Subagents use the same model as the main session, and they can't create subagents.

• BIt won't work as planned, because subagents can't create other subagents

  Correct

  Subagents use the same tools and model as the main session but cannot create other subagents.

• CIt works, but each nested subagent pauses for the developer's feedback

  Incorrect

  Subagents operate without pausing for user feedback, and nesting isn't supported.

• DIt works only in Visual Studio, where nested subagents are supported

  Incorrect

  Subagents aren't currently available in Visual Studio.

In your IDE you can delegate a self-contained subtask to a subagent with its own context window. It works independently without pausing for user feedback and returns its result to the main session. Subagents use the same tools and AI model as the main session, but they cannot create other subagents. They are available in VS Code, JetBrains IDEs, Xcode and Eclipse, not Visual Studio.

• AThe last subagent to finish

  Incorrect

  Subagents focus on their tasks; the orchestrator tracks the goal.

• BThe main agent, acting as orchestrator for its subagents

  Correct

  The orchestrator is usually the main agent that delegates work to subagents.

• CThe rubber duck agent

  Incorrect

  The rubber duck critiques work; it doesn't orchestrate.

• DEach subagent, equally

  Incorrect

  One orchestrator keeps track of the overall goal.

Multi-agent work typically has an orchestrator, usually the main agent, that keeps track of the overall goal. It breaks the goal into tasks, identifies dependencies, assigns each task to a subagent with the context and tools it needs, runs independent tasks in parallel, and reviews and combines the results.

• AEach subagent calls the model independently, adding interactions

  Correct

  Splitting work across subagents can mean more LLM interactions than the main agent alone.

• B/fleet always switches every subagent to the most expensive available model

  Incorrect

  The extra use comes from more model interactions, not a forced model.

• C/fleet bills Actions minutes as credits

  Incorrect

  The extra use comes from model interactions.

• DSubagents re-read the whole repository first

  Incorrect

  The documented reason is independent model interactions per subagent.

Using /fleet can finish independent work faster, keep each agent's context focused and let you assign specialized agents. Because each agent interacts with a model independently, it may also mean more LLM interactions, and therefore more AI credits, than the main agent handling the work alone. Inherently sequential requests gain little.

• Acopilot --acp --fleet "TASK"

  Incorrect

  --fleet isn't supported in ACP server mode.

• Bcopilot /fleet "TASK"

  Incorrect

  /fleet is a slash command inside a session; --fleet is the option.

• Ccopilot --fleet -p "TASK"

  Correct

  --fleet wraps the prompt in fleet mode, and -p runs it non-interactively.

• Dcopilot --cloud -p "TASK"

  Incorrect

  Cloud sandboxes can't be combined with -p, and --cloud isn't fleet mode.

The --fleet option is the command-line equivalent of /fleet. Supply the prompt with -i for an interactive session that starts in fleet mode, with -p to run a task without one, or through standard input. It is especially useful for automation; it isn't supported with --acp.

• AWhenever a task has more than three steps

  Incorrect

  There is no step threshold; an explicit request is needed.

• BAutomatically for every prompt in Autopilot

  Incorrect

  Autopilot doesn't start dynamic workflows by itself; an explicit request is needed.

• CWhenever an extension that defines dynamic workflows is loaded into the session

  Incorrect

  Loading makes workflows available; running one still needs a request.

• DOnly when you ask explicitly, or a skill or slash command tells it to

  Correct

  The chat agent only creates or runs a dynamic workflow on an explicit request or instruction.

Dynamic workflows are defined in Copilot extensions and run on request; you can ask Copilot to run one by name in natural language. Your chat agent only creates or runs a dynamic workflow when you explicitly ask it to, or when a skill or slash command instructs it to.