GitHub Copilot CLI now runs parallel subagents via a new slash command called /fleet. Instead of processing tasks sequentially, /fleet invokes a background orchestrator that decomposes your prompt into discrete work items, identifies which are independent, and dispatches multiple agents simultaneously across different files and directories. Each subagent gets its own context window and shares the same filesystem. The orchestrator polls for completion, resolves dependencies, and synthesizes final outputs.
Prompt quality determines parallelization quality. Vague objectives like '/fleet Build the documentation' collapse into sequential execution because the orchestrator cannot identify independent units. Structured prompts with explicit file boundaries, dependency declarations, and validation criteria like lint or type checks unlock true parallel dispatch. You can also define specialized agents in .github/agents/ with their own models and toolsets, for example routing docs tasks to claude-sonnet-4 while using a lighter model for boilerplate, then reference them directly in your /fleet prompt.
The full article is worth reading for the dependency declaration syntax, the /tasks command that lets you inspect live background agent activity, and the pitfall section that covers filesystem partitioning failures the summary omits entirely. Non-interactive terminal use is also documented, requiring the --no-ask-user flag since subagents cannot respond to prompts mid-execution.
[READ ORIGINAL →]