GitHub Copilot dynamic workflows put agents in code

GitHub shipped dynamic workflows in Copilot CLI, the GitHub Copilot app and the GitHub Copilot SDK. A dynamic workflow is a program defining how a task gets carried out, combining automated steps with one or more agents, where steps run sequentially, in parallel, or both.

The distinction GitHub draws from its own /fleet feature is the useful bit: fleet delegates work to subagents and coordinates them in parallel, whereas a dynamic workflow carries out a process defined in code. You decide the steps, when to involve an agent, and how to use the results. Agents handle the analysis and judgment parts; the orchestration is yours.

The program lives inside a GitHub Copilot extension, which gives it access to Copilot’s extensibility APIs. Capabilities include running commands and calling services, splitting a goal into independent parallel tasks, passing structured results between stages, having subagents verify each other’s findings, asking the user for input where the client supports it, and pausing at a checkpoint to review and resume.

Named use cases are incident investigation, where parallel agents collect logs and telemetry per system and combine into a timeline and root cause report; release checks with human review gates; parallel review of many changed files in a pull request; and finding unresolved review comments on merged pull requests, where the code only reports a finding when two models agree. Workflows can be hand-authored or written by Copilot, and built-in authoring guidance is included.

Availability is broad, all Copilot plans, and in the Copilot app they need no setup. In Copilot CLI they require experimental features, enabled by launching with --experimental or running /experimental on. It is a public preview and subject to change. You can ask “What dynamic workflows are available?” to list existing ones.

(Original source: GitHub)