There’s a category of task that breaks normal agent sessions: the same operation applied across hundreds of items. Audit 300 video transcripts against a ruleset. Triage a backlog of support tickets. Run one transform over every file in a sprawling tree. Hand that to a single Claude Code session and it grinds through serially; reach for subagents manually and you’re hand-wiring the fan-out.
Dynamic workflows are the answer to that shape, and per Leon van Zyl’s walkthrough they’re now generally available. The move: Claude Code writes its own orchestration script, then runs your one prompt across many parallel subagents in a single session. In his demo, 351 agents run at once, each applying the same rules to a different item, results merged into one final output at the end. Crucially it’s not just for code — emails, tickets, transcripts, any batch of similar items qualifies.
How you trigger it is explicit, not magic. You tell Claude Code to use a workflow — van Zyl’s phrasing is to open with Ultra Code then use a workflow for the following, then describe the job. The agent decides what fans out, what verifies, what merges.
The honest caveat — and he leads with it — is cost. Each subagent is effectively its own Claude Code session with its own system prompt, context, and tool calls. Multiply that by hundreds and a task you could have done in one session becomes dramatically more expensive run as a workflow. His rule of thumb: if you know the work fits in a single session, do it in a single session. Workflows earn their price only when the fan-out is real — many independent items, or a job too wide for one context to hold.
The mental model that helps: dynamic workflows are for breadth, not difficulty. A hard single problem doesn’t want 300 agents; 300 easy-but-numerous problems do. Match the tool to the shape of the work, not the size of the ambition.
Source: Leon van Zyl — Claude Code Dynamic Workflows Explained for Beginners (YouTube).