Claude Code can run hundreds of AI agents in parallel on one task. Here’s how to turn it on, when it’s worth the usage and the workflows I’d run first.
You describe a big task in plain English. Claude writes its own plan, spins up as many subagents as the job needs to work in parallel, and has other agents try to poke holes in the results before anything reaches you. You don’t write the plan or manage the agents. You read the final answer.
The example everyone shares: Jarred Sumner used it to port Bun from Zig to Rust. That came to about 750,000 lines of Rust, eleven days from first commit to merge, with 99.8% of the existing test suite passing. (Anthropic: Introducing dynamic workflows)
01
Just ask
In any Claude Code session, say "create a workflow to…" and describe the task. The first time a workflow triggers, Claude shows you what’s about to run and asks you to confirm.
02
Or turn on ultracode
Open the effort menu and turn on ultracode. It sets effort to extra high and lets Claude decide on its own when a task is big enough for a workflow. Anthropic also recommends turning on auto mode for the smoothest runs.
03
Check your plan
Dynamic workflows are now generally available on Pro, Max, Ultra, Team and Enterprise. They’re on by default everywhere except Pro. On Pro, type /config and switch them on. Enterprise admins can turn them off in managed settings.
04
Where it works
The Claude Code CLI, the Desktop app and the VS Code extension, plus the Claude API, Amazon Bedrock, Vertex AI and Microsoft Foundry. If a run gets interrupted, progress is saved, so it picks up where it left off.
Start Small
Anthropic says workflows can use "substantially more tokens than a typical Claude Code session." Make your first run a small, scoped task and watch how much it uses before you point it at anything big.
A workflow earns its usage when the job splits into many similar pieces, you can say what "done" looks like in a checkable way, and the result matters enough to pay for.
05
Personalized cold outreach
Give it a lead list. One agent per lead reads their public site and recent posts, then writes an email that opens with something real about their business. Verifiers check every hook against its source and kill repeated phrases. You get drafts to review and send yourself.
06
Competitor sweep
Drop in your competitor URLs. Agents audit each one in parallel for pricing, positioning and review complaints, verifiers re-check every price and quote, and the final report shows the gaps nobody’s filling.
07
Customer feedback synthesis
Load 90 days of DMs, call transcripts, reviews and support tickets. The workflow splits the pile across agents, then returns the top 5 things to fix, the top 5 marketing angles and verbatim customer language. It’s the proprietary-data moat on autopilot.
08
Site-wide audit
Have it check every page or blog post for SEO, brand voice and anything that blocks a sale. Verifiers grade severity, and you get one report with rewrites ranked by impact.
09
Back-catalog repurposing
Feed it your last 10 podcasts or long videos. Each agent turns one into short posts in your voice, and verifiers make sure nothing repeats across the batch.
Verification is what makes this better than running a bunch of prompts at once. One Claude on a big task makes plausible-sounding mistakes. A hundred Claudes make a hundred of them unless something checks. In a workflow, other agents try to refute each result, and the run keeps going until the answers agree.
You get better results when you spell out what checked looks like for your task. Some examples I use:
New to Claude Code? Start with my CLAUDE.md starter template so every agent knows how your project works.