DeskTheory is where founder-CEOs learn to run their companies on AI leverage.

What are dynamic workflows in Claude Code?

Claude Code's new ability to plan a big job, run tens to hundreds of agents on its pieces in parallel, check the work, and hand you one finished result. Anthropic shipped it on May 28, 2026.

An engineer named Jarred Sumner pointed Claude Code at a 750,000-line codebase and asked it to rewrite the whole thing from one programming language to another. Eleven days later it was done, with 99.8% of the tests still passing. The feature that let a job that big fit in one run now has a name: dynamic workflows. Claude Code can now take on the whole job, not one task at a time.

What it is (in plain English)

Claude Code is Anthropic's tool that runs on your computer and does real work in plain English: research, writing, reading through your files, not only code. Until now it worked one task at a time. You gave it a job, it thought through it in a single line, you watched.

Dynamic workflows change that shape. When you hand Claude a big job, it writes its own plan, splits the work into pieces, and spins up a swarm of smaller copies of itself (subagents, which are separate Claude instances) that each take a piece and work at the same time. Tens, sometimes hundreds, in one session. Before anything reaches you, separate agents check the work, and a piece that does not hold up gets redone. If the run is interrupted, it saves its place and picks back up.

You turn it on two ways: ask Claude to "create a workflow," or switch on the ultracode setting in the effort menu, which lets Claude decide for itself when a job is big enough to deserve one. It is on by default on the Max and Team plans, an admin enables it on Enterprise, and it works in the Claude Code terminal, the desktop app, VS Code, and through the API.

Why you should care as a CEO

For most of its life Claude Code had a ceiling: one task per sitting. Dynamic workflows raise that ceiling to the whole job. The migration that was a multi-week project. The security review you would hire a vendor for. The cross-checked research memo that used to eat a Saturday of open tabs. The unit of work that fits in a single session just changed shape, and Anthropic's launch customers felt it first: Klarna said it surfaces problems "across large codebases" that older tools missed, and CyberAgent called it the gap between "firing off a single subagent and building out a full agent team."

Two things to know before you lean on it. It is a research preview, so expect rough edges. And it spends a lot more than a normal session, because it is running many agents at once, so it asks you to confirm before it starts. Point it at a job where the worst case is "redo it," not "explain it to the board."

Where you'll see it

• The news version, with the launch details: Claude Code can now do a multi-day job in one window.
• The CEO's case for when it's worth the spend, with the costs and caveats laid out: making AI argue with itself before it answers you.
• It is the heavy end of your harness, the setup that lets an agent run real work for you.
• When to reach for this versus a goal: a goal or a dynamic workflow in Claude Code.
• For the everyday version, where you just need to get smart on a topic fast: get up to speed on any topic.

What you should do next

If you are on a Max or Team plan, open Claude Code, point it at the biggest job sitting on your team's backlog, and ask it to create a workflow. If you are not sure whether your job wants a workflow or a goal, read that next.