Claude Code Dynamic Workflows

What Claude Code dynamic workflows are, how ultracode works, when to use workflows instead of subagents or skills, and how to control cost and risk.

Claude Code dynamic workflows let Claude write and run a workflow script that coordinates many subagents in the background. They are for tasks that are too large or too cross-checked for one conversation turn: repository-wide audits, large migrations, security sweeps, broad research, or plans that need independent review before implementation.

Last checked on May 29, 2026. Dynamic workflows are a research preview. Current Claude Code docs say they require Claude Code 2.1.154 or later and are available on paid plans, Anthropic API access, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Availability, cost, admin controls, and limits can change.

Quick Answer

Use a dynamic workflow when the task needs many independent agents, repeatable orchestration, and a single checked report at the end. Do not use it for routine edits, short Q&A, or anything that needs frequent user decisions in the middle.

Start small:

Run a workflow to audit the auth routes for missing permission checks. Report only confirmed findings with file paths, evidence, and the smallest fix.

Then watch the run:

/workflows

If you want Claude to decide when a workflow is warranted during the current session, use:

/effort ultracode

Drop back when the work becomes ordinary:

/effort high

For the dedicated effort setting, read Claude Code Ultracode. If you are here because Opus 4.8 suddenly feels slow or expensive, start with Opus 4.8 slow or expensive troubleshooting.

What Dynamic Workflows Add

Subagents already let Claude delegate focused work. Dynamic workflows move the orchestration itself into a script. That changes the scale and repeatability:

Mechanism	Best for	Limitation
Subagent	One focused side task with isolated context.	Claude still coordinates turn by turn.
Skill or slash command	A reusable instruction pattern in the main session.	The plan still lives in conversation context.
Agent team	Several named workers with team-style collaboration.	Useful, but heavier to configure for one-off research.
Dynamic workflow	Dozens to hundreds of coordinated agent tasks with background progress and cross-checking.	Higher usage, less mid-run interaction, and more operational complexity.

The practical difference is that a workflow can keep intermediate state outside the main chat context, compare independent agent results, restart pieces, and return one coordinated answer instead of flooding the session with every search, test, and file read.

Launch-Window Questions Worth Answering

The Opus 4.8 launch created several related long-tail questions. They should be answered directly instead of hidden inside a generic changelog paragraph:

Question people ask	Direct answer
Is ultracode the same as dynamic workflows?	No. Ultracode is the effort setting; a dynamic workflow is the orchestration run.
Can workflow agents edit code?	They can support coding work, but production edits should be scoped, reviewed, and tested.
Is this only for audits?	No. Good fits include audits, migrations, deprecation sweeps, research, and verification.
Why do I not see workflows or ultracode?	Check Claude Code version, `/config`, plan/provider, and organization settings.
Are workflows always better than subagents?	No. Small noisy side tasks still fit ordinary subagents.
Why is it slow?	Higher effort, background agents, and verification loops can all add time.
Will it cost more?	It can. The official launch notes warn that workflows can consume substantially more usage.

Requirements And Availability

Before relying on dynamic workflows, check four things:

Check	Command or place	Why
Claude Code version	`claude --version`	Official docs require `2.1.154` or later.
Current model and effort	`/model`, `/effort`	Opus 4.8 defaults to high effort; harder workflow runs may use more tokens.
Account or provider	Claude subscription, API, Bedrock, Vertex, or Foundry	Availability depends on plan/provider and organization settings.
Admin settings	`/config` or managed settings	Organizations can disable workflows. Pro users may need to enable the Dynamic workflows row in `/config`.

If you just updated for Opus 4.8 and see thinking-block API errors, update to at least 2.1.156, then restart the affected session if needed.

How To Start A Workflow

There are three common entry points.

Ask For A Workflow

Use the word workflow in the prompt:

Run a workflow to compare every page in content/pages against the sitemap rules and report missing indexable pages.

Claude Code should highlight the trigger and show an approval step before the workflow runs, depending on your permission mode.

Use Ultracode

ultracode combines xhigh effort with automatic workflow orchestration:

/effort ultracode

Use it for a serious session, not as a permanent default. A single request can become several workflow runs: one to understand the repository, one to implement, and one to verify.

Run A Bundled Or Saved Workflow

Claude Code includes /deep-research as a bundled workflow when web search is available:

/deep-research What changed in Claude Code 2.1.154 and 2.1.156?

Saved workflows can also become slash commands. Project workflows live under .claude/workflows/; personal workflows live under ~/.claude/workflows/.

When To Use Dynamic Workflows

Good fits:

Task	Why a workflow helps
Codebase-wide security review	Agents can inspect separate modules and cross-check findings.
Large migration	The workflow can split files or packages into independent batches, then run verification loops.
Dependency or API deprecation audit	Agents can search different patterns and reconcile duplicate findings.
Content or SEO inventory	Agents can compare content files, routes, sitemap rules, and public URLs.
High-stakes plan review	Several agents can draft and challenge the same plan before you choose.

Poor fits:

Task	Better option
Small one-file edit	Normal Claude Code session.
Prompt you repeat often but run in the main context	Skill or slash command.
Task that needs a human decision every few minutes	Plan Mode or a staged manual workflow.
Work that could edit the same files from many agents	Worktrees, narrow stages, or a human-reviewed plan first.
Cost-sensitive long run	Lower effort, smaller scope, or standard subagents.

Cost And Risk Controls

Dynamic workflows can use much more usage than a normal Claude Code session. Treat them like a controlled engineering operation:

Scope the first run tightly.
Ask for read-only analysis before edits.
Tell Claude which directories are in scope and out of scope.
Require evidence for every finding.
Prefer separate workflows for analysis, implementation, and verification.
Check /workflows while a run is active.
Stop the run if it drifts or consumes too much usage.
Save a workflow only after it succeeds on a real task.

For production repositories, combine workflows with Plan Mode, tools allowlist, and hooks. A workflow can coordinate agents; it is not a substitute for repository policy.

Dynamic Workflows And Opus 4.8

Opus 4.8 is the model launch that made dynamic workflows the obvious hot topic, but they are not the same thing.

Topic	What to remember
Opus 4.8	A model update with stronger long-horizon coding behavior, high effort default, and fast mode support.
Claude Code `2.1.154`	The CLI update that introduced Opus 4.8 support and dynamic workflows.
Claude Code `2.1.156`	The follow-up fix for an Opus 4.8 thinking-block API error.
Dynamic workflows	A Claude Code orchestration feature that can run many subagents behind one task.
Ultracode	A Claude Code effort setting that lets Claude decide when to use workflows.

If you are only trying to check or update the CLI version, use Claude Code Latest Version. If you are trying to decide whether a large task deserves background orchestration, stay on this page.

Common Mistakes

Mistake	Why it hurts	Better approach
Turning on ultracode for everything	Usage can climb quickly.	Use it for difficult sessions, then return to high effort.
Starting with a broad migration prompt	The workflow may fan out too much.	Start with inventory or a narrow pilot.
Letting many agents edit shared files	Conflicts and duplicated changes.	Split by module or use worktree isolation.
Treating the final report as verified truth	Agents can still miss context or overstate findings.	Require evidence and run tests.
Saving the first workflow immediately	You may preserve a bad process.	Run it successfully a few times first.
Ignoring admin and provider limits	The feature may not be enabled everywhere.	Check `/config`, plan, and provider documentation.