"Give the agents autonomy" is the agent-tooling industry's mantra. On October 1 GitHub conceded the opposite: explicit control flow, not a trusted delegate.
Dynamic workflows are now in public preview — programs that define in code (not a line of YAML) how a task gets executed: automated steps plus the work of one or more agents, sequential or parallel. Inside: running commands and calling services, structured result passing between stages, one agent verifying another's output, user input prompts, and checkpoints with resumption. Three surfaces: Copilot CLI, the Copilot app, and the Copilot SDK; all plans; enabled with the --experimental flag or /experimental on. The key difference from /fleet, where Copilot hands tasks to subagents on its own: a dynamic workflow is coded ahead of time, for predictability and observability.
Architecture has seen this film before. Choreography versus orchestration: microservices began with free-flowing events, and complex processes ended up under orchestrators with explicit control flow. The plot has replayed with LLM agents in place of services — only faster. A thesis worth arguing with: "agent autonomy" was never the destination, just a temporary artifact of immature tooling. While the flow couldn't be coded, all you could do was trust and hope. The deciding argument for orchestration is debuggability. When free delegation goes wrong, you can't tell which step lost the plot; in a coded workflow every stage and every handoff is an explicit control point. An executor's freedom is a debt, and it's paid by whoever ends up dissecting the failure.
What this means for a team lead. First, split agent tasks by the cost of a mistake. Exploration, drafts, one-offs: free-delegation territory, where mistakes are cheap. Release processes, migrations, anything repeatable — coded workflows only. Second, design checkpoints as rollback boundaries: place them where output can be verified and execution resumed, not mechanically after every step. Third, make stage handoffs a structured contract, not a free-form retelling: one agent can only verify another against a checkable format. And last: this is a preview and the API will shift, so move your processes onto it but keep a way back. Measuring the agents themselves is Cognitus's craft; the architect's job is to give them boundaries.
📎 Source: github.blog/changelog/2026-10-01-dynamic-workflows-in-copilot-cli-and-the-copilot-app