Master Claude Code Dynamic Workflows (49 Agents)
Dynamic workflows are one of Claude Code's most advanced and powerful features, and almost nobody is using them. Instead of handing one big task to a single agent, Claude builds a whole team of specialized subagents, each with its own model and its own tools, and fans the work out across them. In this walkthrough I explain how they differ from static workflows, then trigger one live: a market study across France and the Netherlands that spun up 49 agents and burned through 4 million tokens.
What you need
- Claude Code. Dynamic workflows are only available there right now; on Cowork you can only use static workflows
- The $100 to $200 Claude plan, because these runs consume a lot of tokens
- A long, complex task that genuinely needs to be fanned out
1. Static vs dynamic workflows (0:10)
Everything starts from your prompt for a long, complex task. Claude then selects whether to use a static workflow or a dynamic one.
The key difference: in a static workflow, the team of subagents is made of general purpose agents, not specialized at anything. In a dynamic workflow, Claude actually selects the agent type and the tools for each subagent. The output you get is truly specialized to your use case. Think analyzing a business plan, or changing the entire UI of a website at once. Both paths end in a summarized result, and even static workflows get you quite far ahead of most people, but dynamic workflows give you something that really suits the job.
2. Each agent gets its own model and its own tools (1:06)
Once the workflow is running, Claude breaks it down into tasks and hands them to specialized subagents. For agents that do not need a lot of brainpower, it selects small models. For the ones that do, it selects bigger models.
It also right-sizes the tools. If you use Google Drive on your main Claude account and the subagents do not need that access, it just takes it away. That reduces the risk of hallucinations and improves the quality of the output. A nice bonus: the agents talk between themselves, which gives you a better quality of response overall.
3. Building one: the market study prompt (2:36)
Time to trigger our own. We ask Claude Code to help with searching information on the web, and we literally tell it to use a dynamic workflow. Here is the exact prompt from the video:
Hey Claude, I am running a market study. I need to break it down into a dynamic workflow. First, I need a team of agents to analyze country data, and then a second specialized team to analyze company data. We're targeting France and also the Netherlands, and I want to study companies that are doing heavy investments in AI. At most five companies per geography. And here comes the key part: please use a dynamic workflow for this.
You do not even need the extra reasoning tier for this; standard high reasoning is good enough. It starts a session and asks a few questions. You can also combine this with more advanced features, like looping your agents and having subagents call other subagents.
4. The two specialized teams it created (4:14)
It took about three minutes for Claude Code to set up all of the specialized agents. It gave me the breakdown of the two teams: team one is the country specialists, team two is the company specialists that run on the shortlisted companies.
It even puts out a card you can click to watch your agents work: the tokens they have generated, the tools they have used, how long they have been running. You become your own little manager for this team. Within literally fifty seconds of running, we were at four hundred thousand tokens generated.
5. Token usage and which plan you need (4:57)
You have to be extremely careful here, because this consumes a lot of tokens. You must have the $100 to $200 plan for Claude Code, or this will eat all of your tokens very fast.
My suggestion: on the $100 plan, do this at most once or twice per day, and that is already pushing it. On the $200 plan you can run it a couple of times per day with no problem.
6. Why this task needs a dynamic workflow (5:40)
Claude Code actually explains why a dynamic workflow is relevant for this use case. Since we do not know in advance which companies we will be researching, the first team of agents passes its findings down to the second team, and the second team produces the report I actually want.
That dependency is the tell. When one wave of research decides what the next wave works on, a single agent or a static team will not cut it. This is also why you should use dynamic workflows for big jobs like digging through your CRM or marketing campaigns: dumping a pile of information on one single instance of Claude diminishes the quality of the output significantly.
7. The result: 49 agents and 4 million tokens (10:13)
The final tally: a whopping 49 different subagents, each specialized per company with its own methodology for researching information. It consumed 4 million tokens, literally half of all the tokens I could use for the next five hours. But the output quality is one hundred percent there.
It did not only search the internet. It also used the tools available inside my workspace to find the most up to date information. The deliverable is a lastmarketstudy.md file, plain Markdown, with a brief overview. From there I can ask it to generate a slide deck or just hand me the research brief.
8. Best use cases (11:29)
Map out the use cases that come to mind and start testing. Dynamic workflows are fantastic for recruitment, for product development, for checking your CRM for reactivation, and for analyzing entire datasets, Excel files or Google Sheets with thousands and thousands of rows.
Pitfalls to avoid
- Running this on a small plan. Without the $100 to $200 tier, one run can wipe your usage.
- Overusing it. Even on the $100 plan, once or twice a day is the ceiling.
- Using it for simple tasks. If the work does not need fanning out, a single agent is cheaper and just as good.
Where to go next
Pair this with the command that keeps a session running until the job is done: Make Claude Code Work for Hours. And when the research is finished, share it properly with Claude Code Artifacts Explained in 3 Minutes.
