Dynamic workflows

Orchestrating many subagents from a script Claude writes — the ultracode keyword, /workflows, saving runs as commands, limits, resume, and cost.

A dynamic workflow is a JavaScript script that orchestrates subagents at scale. You describe the task, Claude writes the script, and a runtime executes it in the background while your session stays responsive.

The point isn’t just “more agents”. It’s that the plan moves into code: the loop, the branching, and the intermediate results live in script variables instead of in a context window. Claude’s context only holds the final answer.

When it’s worth it

Subagents, skills, agent teams, and workflows can all run a multi-step task. The difference is who holds the plan:

Reach for a workflow when the task is bigger than one conversation can coordinate, or when the same step has to run across many items: a codebase-wide bug sweep, a 500-file migration, a research question whose sources need cross-checking, a hard plan worth drafting from several angles first.

Width or depth

The feature workflows are most often confused with isn’t subagents — it’s /goal, because both let Claude run a long way without you. They point in different directions:

/goal all tests pass and the linter is clean keeps working until that’s true. A workflow fans a hundred agents across a hundred files once, then synthesizes. Both punish a vague brief — one by never converging, the other by spending in parallel.

They compose (a workflow inside a goal loop), which is powerful and an excellent way to spend a lot of money quickly. Bound the scope and name the deliverable before either.

Starting one

ultracode: audit every API endpoint under src/routes/ for missing auth checks

Claude Code highlights the keyword in your input and writes a script instead of working turn by turn. The keyword only changes how the work is structured — the run stays inside the session’s permission mode and sandboxing.

The bundled one

/deep-research <question> ships with Claude Code and is the fastest way to see a workflow run. It fans out web searches across several angles, cross-checks the sources it finds, votes on each claim, and returns a cited report with the claims that didn’t survive filtered out. It needs the WebSearch tool available, and it only runs when you invoke it.

Approval

The per-run prompt lists the planned phases with Yes, run it, Yes, and don’t ask again for this workflow in this project, View raw script (Ctrl+G opens it in your editor), and No. Tab lets you adjust the prompt before the run starts.

What the script looks like

Plain JavaScript with top-level await: a meta block, then a body that spawns agents.

export const meta = {
  name: 'audit-routes',
  description: 'Audit every route handler for missing auth checks',
}

const found = await agent('List every .ts file under src/routes/.', {
  schema: { type: 'object', required: ['files'], properties: { files: { type: 'array', items: { type: 'string' } } } },
})

const audits = await pipeline(found.files, file =>
  agent(`Audit ${file} for missing authentication checks.`, { label: file }),
)

return audits.filter(Boolean)

Watching and managing a run

/workflows lists running and finished workflows; select one and press Enter for its progress view, which shows each phase with agent counts, token totals, and elapsed time. A one-line summary also appears in the task panel below the input box (↓ to focus, Enter to expand).

Saving a run as a command

Press s in /workflows, then Tab to pick where it lands:

Project scripts are shared with everyone who clones the repo; the home directory ones follow you across projects and are yours alone. Either way the workflow becomes /<name> in future sessions and shows up in / autocomplete. If both locations define the same name, the project one wins. Workflows shipped in a plugin live in a workflows/ directory at the plugin root and run namespaced — /acme-tools:release-audit.

Saved workflows take input through args:

Run /triage-issues on issues 1024, 1025, and 1030

Resuming

A stopped run can be resumed with p from /workflows, within the same session — exiting Claude Code loses it. Two rules decide what’s kept:

That’s why a fan-out of many small agents preserves far more progress than one long-running agent.

Limits and cost

A run can burn meaningfully more tokens than doing the same task in conversation, and it counts against your plan’s limits like any other session. Claude Code flags a run scheduling more than 25 agents or projected past 1.5M tokens with a Large workflow warning in the task panel — advisory only, and hidden when ultracode is on.

Size guideline

Advice to Claude about how many agents to aim for, not a cap. Set it with /config (Dynamic workflow size), /config workflowSizeGuideline=small, or the workflowSizeGuideline settings key.

Turning it off

Toggle Dynamic workflows off in /config, set "disableWorkflows": true in ~/.claude/settings.json (or in managed settings for a whole org), or set CLAUDE_CODE_DISABLE_WORKFLOWS=1. The bundled commands disappear, ultracode stops triggering runs, and it leaves the /effort menu.