GitHub Copilot adds dynamic workflows that let a team orchestrate multiple agents from code

Writing a Copilot prompt and clicking run leaves the agent in charge of how the task gets done, which is the right trade for a quick change and the wrong one for a release check that runs every Friday. GitHub shipped dynamic workflows today across Copilot CLI, the Copilot app and the Copilot SDK: a program that defines the steps itself and calls one or more agents only where it needs analysis or judgment.

The announcement is a changelog post dated 1 October 2026. Dynamic workflows are in public preview on all Copilot plans.

What a dynamic workflow is

A dynamic workflow is code that defines how a task gets done. It can run commands, call tools, divide a goal into tasks that run in parallel, pass structured results from one stage to the next, pause at a checkpoint for a human review and resume on command, and ask for input where the client supports it. The program lives inside a GitHub Copilot extension and so has access to the same APIs Copilot extensions use.

The worked example in the changelog: investigate a service incident. The workflow collects logs and telemetry, assigns independent agents to analyse different systems in parallel, and combines the agents' structured findings into a timeline and a root-cause report. The steps run the same way every time. The agents only decide what the structured output says; the code decides what happens with it.

GitHub lists five other shapes it expects to be good fits. Running release checks and asking one agent to assess the failures before a human reviews. Reviewing many changed files in parallel on a pull request. Sweeping a large codebase for a pattern (missing tests, use of an API that is being removed) across many directories at once. Asking two models whether a merged pull request's unresolved review comment still matters and only reporting findings when both agree. Researching a change, planning its implementation from the findings and then making the change.

How it differs from /fleet

Copilot already has /fleet, which delegates work to subagents and lets Copilot coordinate them. A dynamic workflow carries out a process defined in code. The practical difference: in /fleet the model picks each next step, so two runs on the same input can take different paths; in a dynamic workflow the path is the code, so the same inputs produce the same path.

How to use it

Dynamic workflows are always available in the GitHub Copilot app with no setup. In Copilot CLI they are behind an experimental flag. GitHub says Copilot itself will write a workflow for you if asked, with built-in authoring guidance; the step-by-step is in the Copilot docs under "Creating a dynamic workflow". Feedback goes through /feedback in the CLI, and the discussion thread is on GitHub Community.

The announcement does not state the pricing cost of a workflow run beyond saying it is available on all Copilot plans, so the per-step cost presumably comes out of the same model budget as any other Copilot call. GitHub's own guidance is to use a workflow when the process is worth defining once and reusing, and to use a normal chat prompt when it is not.

Source

• Primary source: GitHub Changelog: Dynamic workflows in Copilot CLI and the Copilot app, 1 October 2026