Anthropic's dynamic workflows let Claude Code write a script that fans a task out to dozens or hundreds of subagents, then runs it in the background while your session stays free. Turn on ultracode with /effort ultracode and Claude stops waiting for you to ask. It plans a workflow for every substantial task, often several in a row: one to understand the code, one to make the change, one to verify it. This is the most powerful mode the tool has, and the one most likely to flatter you into believing scale is the same thing as progress.

The pricing is honest if you actually read it. Each subagent burns its own tokens, so a hundred agents working a hundred files costs roughly a hundred times what one agent on one file does. Ultracode compounds that by turning a single request into a chain of workflows. Anthropic caps a run at 1,000 agents with 16 running at once, which stops a runaway loop while doing nothing about the cost of a run you fully meant to launch.

The orchestration is real engineering. The plan lives in a JavaScript script the runtime executes on its own, away from the conversation, so intermediate results stay in variables instead of clogging Claude's context. Agents editing files in parallel can each be handed their own git worktree, a private copy of the repo, so two of them never write over the same line. Nothing improvises; the script decides what runs next.

That is the gap the caps don't cover: they protect you from accidents, not from intent. The machinery guarantees the agents won't collide, but it has nothing to say about whether you needed a hundred of them. A weak plan doesn't get better when you run it in parallel; it just gets more expensive to be wrong.

Sources: