Claude Code Dynamic Workflows: Official Step-by-Step Guide

What Launched

On May 28, 2026, Claude introduced dynamic workflows in Claude Code. The official launch post describes them as long-running, parallel runs where Claude writes orchestration code, fans work out to many subagents, and checks results before reporting back. The verified ClaudeDevs thread framed the product behavior more directly: use the word workflow in a prompt to get started.

The useful way to read this launch: workflows are not another prompt template. They are a new orchestration path inside Claude Code for work that is too large for one conversational loop: codebase audits, service-wide bug hunts, large migrations, cross-checked research, and plans that should be attacked from several angles before you accept them.

The workflow shape:

Your prompt
  -> Claude writes workflow JavaScript
  -> Runtime starts background phases
  -> Subagents investigate, edit, test, or review
  -> Results are cross-checked
  -> One coordinated answer returns to the session

Definition

A Claude Code dynamic workflow is a JavaScript orchestration script that Claude writes for your task, then a separate runtime executes in the background.The agents do the actual reading, writing, shell work, research, and review. The script coordinates the run and stores intermediate results outside the main chat context.

That distinction is the whole feature. In normal chat, Claude decides what to do next turn by turn, and every result has to fit back into the conversation. In a workflow, the plan, loops, branches, and intermediate values live in the workflow script. The main session stays responsive and receives the coordinated result.

Requirements and Availability

The operational source of truth is the Claude Code workflow docs. They say dynamic workflows are in research preview, require Claude Code v2.1.154 or later, and are available on paid plans and through supported API/provider deployments. The launch blog and ClaudeDevs thread also list Claude Code CLI, Desktop, VS Code, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

# Check your version
claude --version

# Update if needed
claude update

# Inside Claude Code, open settings and confirm Dynamic workflows
/config

There is one launch-day nuance: the blog and X thread emphasize Max, Team, and Enterprise availability, while the docs currently give the more specific paid-plan setup note and mention Pro enablement through /config. For actual setup, follow the docs page and your account's settings screen.

Mental Model: Subagents, Skills, and Workflows

Claude Code already had several ways to parallelize or reuse work. Dynamic workflows sit above subagents: they use subagents, but the coordination moves into a script. That is why the feature is useful for repeatable fan-out and verification patterns.

Use subagents when you need one or two isolated investigations. Use skills when the know-how should be reusable. Use workflows when the orchestration itself should be repeatable: fan out, compare findings, restart failed agents, save the successful run, and reuse it as a command.

Quickstart: Run the Built-In Workflow

The fastest safe test is /deep-research, the bundled workflow documented by Claude Code. It fans out searches across angles, fetches sources, cross-checks claims, and returns one cited report. Pick a bounded question first so you can see usage and behavior before trying a large migration.

Step 1: Start Claude Code in a project

cd /path/to/your/project
claude

Step 2: Run a scoped research workflow

/deep-research What changed in the Node.js permission model between v20 and v22?

Step 3: Approve the workflow

Claude Code shows the planned workflow before it runs. Choose Yes, run itif the phases match your intent. If you want to inspect the generated orchestration, choose View raw script. If the prompt is too broad, cancel and narrow it.

Step 4: Watch it from /workflows

/workflows

The workflow view shows phases, agent counts, token totals, and elapsed time. Drill into a phase to inspect each agent's prompt, recent tool calls, and result. When the run finishes, Claude inserts the report into your current session.

Create a Custom Workflow With One Prompt

To run your own task as a workflow without changing the whole session, include the word workflow anywhere in the prompt. Claude Code highlights the word and routes the task through workflow creation.

Run a workflow to audit every API endpoint under src/routes for missing auth checks.

Scope:
- read-only first pass
- group findings by endpoint
- verify each finding with an independent reviewer agent
- return only confirmed issues
- do not edit files unless I approve a follow-up workflow

The docs mention one small escape hatch: if Claude Code highlights workflowand you did not mean to trigger the feature, press alt+w to ignore it for that prompt.

Prompt Pattern

Use Ultracode When You Want Claude to Decide

ultracode is not a normal model effort level. The model configuration docs define it as a Claude Code session setting: it uses xhigh reasoning effort and additionally has Claude orchestrate dynamic workflows for substantive tasks.

# Turn on automatic workflow orchestration for this session
/effort ultracode

# Return to normal high effort for routine work
/effort high

Use ultracode for a session where most tasks are large, ambiguous, or high-risk: migrations, deep audits, design stress tests, or repo-wide cleanup. Do not leave it on for routine edits. The workflow docs state that ultracode applies to every substantive task in the session, uses more tokens, takes longer, and resets when you start a new session.

Approval Flow: Read Before You Run

The CLI approval prompt shows the planned phases and gives four actions. Treat this as the last checkpoint before the run can consume meaningful usage or make approved file edits through agents.

Two keyboard details matter. Ctrl+G opens the generated script in your editor. Tab lets you adjust the prompt before the run starts. In Desktop, the same decision appears as an approval card with Once, Always, and Deny actions.

Save a Workflow as a Slash Command

When a workflow produces a useful repeatable process, save it. Saved workflows become slash commands and appear in / autocomplete like bundled commands.

/workflows

# Select the completed run
# Press s
# Choose project or user location
# Press Enter to save

# Later:
/api-auth-audit

If a project workflow and a personal workflow share a name, the project workflow wins. That is the right default for team consistency, but it also means workflow names should be specific.

Monitor, Pause, Resume, Restart, and Stop Runs

Workflows run in the background. Use /workflows to inspect live and completed runs, then use the keys shown in the footer. The task panel below the input also shows a one-line progress summary while a run is active.

Resume is session-scoped. If you pause or stop a run, completed agents can return cached results and the remaining agents can run live when you resume in the same Claude Code session. If you exit Claude Code while a workflow is running, the next session starts the workflow fresh.

Runtime Behavior and Limits

The workflow runtime is isolated from the conversation. The workflow script coordinates the agents; the agents do the file, shell, web, and MCP work. The docs list concrete limits so users can reason about resource usage.

Cost and Usage Control

Both the docs and the ClaudeDevs thread warn that workflows can consume far more usage than a normal Claude Code conversation. A workflow can start many agents, and each agent uses your session's model unless the script routes a phase to another model.

# Before a large workflow
/model
/usage

# In your prompt
Use a smaller model for indexing and low-risk classification phases.
Use the strongest model only for final verification and architectural judgment.

• Start with a scoped run, not the entire monorepo.
• Check /model before launching expensive work.
• Use /usage to inspect current session usage and plan-limit impact.
• Ask Claude to route low-risk stages to a smaller model when appropriate.
• Stop runaway or mis-scoped runs from /workflows.
• Keep ultracode for sessions where most tasks justify workflow orchestration.

Permissions and Safety Details

Your permission mode controls the workflow launch prompt, but not every subagent action behaves like your main turn. The docs state that workflow subagents run in acceptEdits mode and inherit your tool allowlist. File edits are auto-approved. Shell commands, web fetches, and MCP tools outside the allowlist can still prompt mid-run.

For high-risk work, use a two-stage pattern: first workflow produces confirmed findings only; second workflow applies the approved changes. That matches the docs' note that workflows do not accept arbitrary mid-run user input.

Turn Workflows Off

Dynamic workflows can be disabled at the user or organization level. When disabled, bundled workflow commands are unavailable, the workflow keyword no longer triggers a run, and ultracode is removed from the /effortmenu.

# Option 1: interactive setting
/config
# Toggle Dynamic workflows off

# Option 2: user settings
{
  "disableWorkflows": true
}

# Option 3: environment variable, read at startup
CLAUDE_CODE_DISABLE_WORKFLOWS=1 claude

Organization admins can also set "disableWorkflows": true in managed settings or use the Claude Code admin settings page. Use that for environments where multi-agent token usage or auto-approved file edits need central policy.

Workflow Examples You Can Adapt

Anthropic's launch post lists codebase-wide bug hunts, profiler-guided optimization, security audits, migrations, modernization work, and high-stakes double-checking as early workflow uses. The same post cites a Bun port from Zig to Rust as an at-scale example, with the caveat that it was not in production at publication time.

Example 1: API Auth Audit

Run a workflow to audit src/routes and src/api for missing authentication checks.

Rules:
- read files only
- group endpoints by auth pattern
- have one agent find issues and another verify them
- exclude test fixtures and generated files
- return confirmed findings with file paths and suggested fixes

Example 2: Migration Plan Before Edits

Run a workflow to plan the migration from the old payments client to the new billing SDK.

Do not edit files.
Map every call site, classify risk, propose phases, and have reviewer agents challenge the plan.
Return a phased migration plan with test commands for each phase.

Example 3: Cross-Checked Research

/deep-research Compare the official migration paths for React 19 Server Actions and Next.js 16 forms.

Focus on official docs.
Return only claims that are supported by source links.

Example 4: Save a Team Review Command

Run a workflow to review this branch for correctness bugs.

Scope:
- current git diff only
- no style-only feedback
- verify each finding against code
- include test commands that would catch the bug

# After the run succeeds:
/workflows
# press s and save to .claude/workflows/branch-correctness-review

Common Mistakes

Launch Checklist

• Confirm claude --version is at least v2.1.154.
• Open /config and confirm Dynamic workflows are enabled.
• Run a small /deep-research workflow before code edits.
• Use /workflows to inspect phases, token totals, and agent output.
• For custom workflows, include path scope, output format, verification rule, and edit policy.
• Use /effort ultracode only for sessions where most tasks justify it.
• Save repeatable workflows to .claude/workflows/ for team commands.
• Use /usage, /model, and smaller-model routing for cost control.
• Disable workflows through /config, settings, env vars, or managed settings when policy requires it.

FAQ

Glossary

Official Sources

This post intentionally uses only official Claude, Claude Code, Anthropic-owned, and verified ClaudeDevs sources.

• Claude Blog: Introducing dynamic workflows in Claude Code
• Claude Code Docs: Orchestrate subagents at scale with dynamic workflows
• ClaudeDevs official X thread: Dynamic workflows launch
• Claude Code Docs: Commands
• Claude Code Docs: Run agents in parallel
• Claude Code Docs: Model configuration and effort levels
• Claude Code Docs: Settings
• Claude Code Docs: Manage costs effectively
• Claude Code Docs: Quickstart

Related: the Opus 4.8 official launch breakdown and the Opus 4.8 API migration guide.