WORKFLOWS
Dynamic Workflows Cheat Sheet 🚀
Claude Code can now run hundreds of AI agents in parallel on one task. Here's how to turn it on, when it's worth the tokens, and the workflows I'd actually run.
Source: Anthropic: Introducing Dynamic Workflows →
Dynamic workflows is the biggest Claude Code feature drop of 2026 so far. You describe a big task in plain English, Claude writes its own orchestration plan, spins up tens to hundreds of subagents to do the work in parallel, and other agents check (and try to refute) the results before anything reaches you. You don't write the plan. You don't manage the agents. You read the final answer.
The receipt everyone's sharing: a developer used dynamic workflows to port the Bun runtime from Zig to Rust — roughly 750,000 lines of code, 11 days from first commit to merge, 99.8% of the existing test suite still passing. That used to be months of work for an entire team.
This guide is the cheat sheet: exactly how to turn it on, when it's the right tool, the 5 real workflows I'd run first, and the cost rules so you don't accidentally burn a month of your plan in an afternoon.
How To Actually Turn It On
There are two ways in, and they do the same thing under the hood:
· Method 1 — Just ask Claude. In any Claude Code session on a supported plan, say "create a workflow to…" followed by your task. Claude writes the orchestration script, shows it to you, and asks you to confirm before any agents run.
· Method 2 — Turn on ultracode. In your effort menu (the same place you set low / medium / high), flip ultracode. This sets your effort to xhigh AND lets Claude decide on its own when a task is big enough to deserve a workflow. Best for set-it-and-forget-it usage.
Where it works: Claude Code CLI, Desktop, and the VS Code extension. Also on the Anthropic API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.
What plan you need: Max, Team, or Enterprise. Not Pro. On Enterprise the feature is off by default and needs an admin to enable. So if you're on Pro and don't see ultracode, that's why — upgrade or wait for general availability.
Before any workflow actually runs: Claude shows you the plan first — what subagents it's about to spin up, what each one will do — and asks you to confirm. No surprise runs.
Start scoped
Anthropic's first official tip: start with a small, scoped task on your first run. Dynamic workflows can consume substantially more tokens than a normal Claude Code session. Watch one go end-to-end before you point it at your codebase.
When To Use A Workflow (And When Not To)
Workflows shine when the task has three properties: parallelizable, verifiable, and worth the tokens. If any one of those is missing, run a normal session.
Use a workflow when:
· The task naturally fans out (audit every file, score every applicant, scrape every competitor).
· You can describe what "done" looks like in a checkable way.
· The answer is worth real money (a migration, a launch audit, a customer-data synthesis you're going to act on).
Skip workflows for:
· Chat-style work or anything you'd ask in a sentence.
· Quick rewrites, short summaries, one-file edits.
· Tasks where you don't know what good looks like yet — figure that out in a regular session first, THEN scale with a workflow.
5 Real Workflows To Run This Week
These are the five workflows that genuinely earn their tokens. Each one would take you (or your team) days the old way.
This is the master prompt I use to kick off a personalized-cold-outreach workflow. Adapt the bracketed inputs to your business and run it.
The Token Cost Rules
Anthropic was upfront about this in the announcement: dynamic workflows consume meaningfully more tokens than a typical Claude Code session. A hundred subagents is a hundred Claudes calling tools, reading files, and verifying each other. So the rules:
· Start with one scoped run before anything big. Watch how it spends. Get a feel.
· Ask Claude for an estimate before confirming. When it shows you the plan, ask "how many tokens do you expect this to use?" Then decide.
· Don't workflow on chat-tier work. If a single Claude can do it in 30 seconds, a workflow is a waste.
· Save the receipts. When a workflow earns its tokens (real ROI), save the plan and prompt as a template so you can rerun without re-thinking.
The Verification Step Is Why This Beats Single Agents
The thing that separates dynamic workflows from "just run a bunch of prompts in parallel" is the verification layer. After a subagent finishes a piece of work, other subagents try to refute what it found before it gets folded into the final answer.
Why this matters: a single Claude on a big task makes plausible-sounding mistakes. A hundred Claudes on a big task make a hundred plausible-sounding mistakes — unless you check. The refute-step is the check.
What to do with this: when you ask for a workflow, explicitly tell Claude what "verified" looks like for your task. "Verify by running the test suite. Verify by checking the output against this rubric. Verify by having a second agent re-read the email and flag any sentence that sounds AI-generated." The more specific your verify rule, the better the final answer.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.