Claude Opus 4.8 introduced Dynamic Workflows — Claude Code's orchestration layer for jobs too big for a single pass. It writes its own plan, runs subagents in parallel, and verifies the output.
A Dynamic Workflow is Claude Code planning and running its own multi-agent job. Instead of you hand-building a chain of steps, Claude writes an orchestration script, spins up subagents to do pieces of the work in parallel, and checks the results before calling it done.
Think of it as a manager that hires its own team for the task in front of it, then disbands them when the job is finished.
This is for work that doesn't fit in a single context window: large refactors across hundreds of files, repo-wide test generation, multi-file security audits, big migrations. If you've ever watched a normal session run out of room, that's the signal.
For small, single-file tasks, a regular session is still faster. Dynamic Workflows shine on breadth.
State the big goal in plain language and let Claude plan. You don't write the orchestration — Claude does. It decides how to split the work, how many subagents to run, and how to recombine the results.
Your job is a clear objective and good guardrails (what done looks like, what not to touch).
> Migrate the codebase to the new API. Keep public behavior identical. Add tests for anything you change.
The workflow verifies its own outputs and hands you a summary: what passed, what needs a look, what failed. You review at that level instead of reading every file, then re-run only the parts that need it.
That's the whole point — the work scales without your attention having to scale with it.
A fill-in-the-blank brief that gives Claude a big goal plus the guardrails it needs to plan a safe workflow.