Claude Code Dynamic Workflows: How to Cap Your Agent Fleet
Claude Code dynamic workflows let one prompt fan out into a fleet of agents working in parallel, and by default nothing tells that fleet how big it is allowed to get. I watched a single web search turn into 116 agents. The control is one row in your settings called Dynamic workflow size, it takes four values, and it ships on the value that sets no ceiling at all. This post covers what each value does, how to change it, and the part the setting does not actually control.
Key Takeaways
- Type
/configin a Claude Code session, find Dynamic workflow size, and pick a value. That is the whole change.- The four values are unrestricted, small, medium, and large. Unrestricted is the default and carries no guideline.
- The tiers are guidance, not a hard stop. Your terminal reads "aim for" for a reason, so treat the number as a target rather than a wall.
- I run large. It keeps a trivial task from turning into a fleet without capping the runs where the fan out is the point.
- A dynamic workflow spends tokens like any other session, so the size you pick is a usage decision before it is a speed decision.
Claude Code moves quickly, so treat the row labels and version numbers below as current at the time of writing.
What Are Claude Code Dynamic Workflows?
Claude Code dynamic workflows are an orchestration layer that breaks one request into many agents running in parallel, instead of working through it in a single pass. Claude writes a script that decides what to run, how much of it runs at once, and what gets checked, then executes it. The feature needs Claude Code v2.1.154 or later, and Anthropic's workflow docs put a ceiling of 16 agents running concurrently and 1,000 agents total on any one run.
That ceiling is the tell. A feature with a thousand agent budget is built for work that deserves a thousand agents. Deep research, a long loop, several rounds of verification against each other. Point the same machinery at a question you could have answered yourself and it still brings the fleet.
The Setting That Caps the Fleet: Dynamic Workflow Size
Dynamic workflow size is the row in Claude Code settings that tells a workflow roughly how many agents to aim for. It takes four values, and Anthropic's docs describe them like this.
| Value | Guideline | Reach for it when |
|---|---|---|
| unrestricted | No guideline. This is the default. | The fan out is the whole point |
| small | Aim for fewer than 5 agents | Everyday work you want kept tight |
| medium | Aim for fewer than 15 agents | A real task with some checking |
| large | Aim for fewer than 50 agents | Research and verification runs |
In the terminal the same values read as small (aim for <5 agents), medium (aim for <15 agents), large (aim for <50 agents), and plain unrestricted with no qualifier after it. The setting itself needs Claude Code v2.1.202 or later, which is newer than dynamic workflows themselves, so a version that runs workflows will not necessarily show you this row.
How to Change Dynamic Workflow Size
Changing Dynamic workflow size takes one command and about ten seconds.
- Open your config. Type
/configinside any Claude Code session. - Find the row. Dynamic workflow size sits below Dynamic workflows, the master toggle, and below Ultracode keyword trigger.
- Read what it says now. On a fresh setup it says unrestricted. That is the default, and it is the reason a small job can bring a large fleet.
- Pick a value. Arrow through unrestricted, small, medium, and large, and select one. It saves as you go.
Nothing else has to change. The master Dynamic workflows toggle stays on, and the size row only governs how wide a workflow spreads once one starts.
Which Size I Run, and Why
I use dynamic workflows constantly, and accuracy is why. When I need what I am about to publish to be factually correct, a workflow that sends several agents to check the same claim from different angles catches things a single pass does not. That is the job the feature is genuinely good at.
It also gets out of hand. I ran a dynamic workflow that was a web search, and it launched over 116 agents. One search. That is the run that made me go find this setting.
I set mine to large. Small at fewer than 5 agents caps the thing I actually want the feature for, and unrestricted is how you get 116 agents out of a search. Large leaves room for the verification runs without letting a quick question turn into a fleet. It is also a setting, not a decision you live with, so I change it when a specific run deserves more room and change it back afterward.
What the Size Guideline Does Not Do
The word on your screen is "aim", and it was chosen carefully. Anthropic's docs say Claude Code sends the value to Claude as advice, so a prompt that calls for a different scale still overrides it. A run can come in above the tier you picked and nothing has gone wrong.
Two hard limits sit underneath, and those do not move. Anthropic's docs put 16 agents as the most that run at once and 1,000 as the most any single run will use in total. They apply whichever value you pick, including unrestricted, so the default is bounded, just very loosely.
The setting does change one threshold. Claude Code shows a large workflow warning when a run schedules more than 25 agents or projects more than 1.5 million tokens, and per the docs, once you set a size guideline the guideline's agent count takes over from that 25 agent trigger. So picking small does not only shrink the fleet, it moves the line at which Claude Code tells you the run is getting big. The warning needs Claude Code v2.1.203 or later, and it is a heads up, not a stop.
Dynamic Workflows and Your Usage Limit
A dynamic workflow counts toward your plan's usage and rate limits like any other session, per Anthropic's docs. Fifty agents is not fifty free agents. That is what makes this a usage setting, and it is why the default costs you the most on exactly the tasks that deserve it least.
Model choice compounds it. If you are running dynamic workflows, run them on Sonnet or Opus rather than Fable, because a fleet on Fable burns through your allowance faster. The gap is not small. At the time of writing, Anthropic's pricing puts Claude Fable 5 at $10 per million input tokens against $5 for Claude Opus 5 and $2 for Claude Sonnet 5. The multiplier is the problem: a model that costs twice as much per token costs twice as much 116 times over. If you are weighing the two mainstream options for a session, Sonnet and Opus split more cleanly than the pricing suggests, and if usage is the constraint you keep hitting, an advisor setup gets you close to top model quality on a cheaper model's budget.