Claude Dynamic Workflows — A Comprehensive Guide: What's New and What It Means for Enterprise AI

Hello, this is Hamamoto from TIMEWELL.

On May 28, 2026, Anthropic released Claude Opus 4.8 and, alongside it, shipped "Dynamic Workflows" as a research preview. Within hours of the announcement, the one-liner making the rounds in engineering communities was "Claude writes its own orchestration script and runs up to 1,000 subagents in parallel." That single sentence captures part of the story, but explaining what this feature actually is to executives and business leaders takes a more careful breakdown.

In this article, I walk through Dynamic Workflows using Anthropic's official blog and product documentation as the primary source. I explain what it is, how it differs from existing primitives like Tool Use, the Agent SDK, and Managed Agents, and what questions you should be asking before bringing it into your enterprise. The intent is to give executives, DX leaders, and engineering managers something they can take into an internal decision-making meeting without further translation.

The information in this article is drawn primarily from Anthropic's official blog post "Introducing dynamic workflows in Claude Code" and the Claude Code documentation page "Orchestrate subagents at scale with dynamic workflows," with supporting coverage from TechCrunch, The New Stack, and VentureBeat. I have not made any claim that rests on reporting alone.

What Is Dynamic Workflows — The Official Definition and the Launch Facts

Let's start with the primary source. Anthropic's documentation defines Dynamic Workflows as follows:

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

In plain English, a dynamic workflow is a JavaScript script that coordinates subagents at scale. Claude writes the script in response to a task you describe, the runtime executes it in the background, and your interactive session stays responsive while the script runs.

For readers newer to this space, a few terms are worth defining:

• Subagent: A "spawned instance" of Claude that handles a specific task. Subagents run in isolated contexts and permission scopes, and they can run in parallel.
• Orchestration: The work of controlling the execution order, branching, and concurrency across multiple agents.
• Runtime: An independent execution environment for the script. Here, it refers to a sandbox separate from Claude's conversational thread.

Launch Date, Eligible Models, and Plans

The launch facts below are stated explicitly in Anthropic's official communications and the Claude Code documentation.

Item	Details
Launch date	May 28, 2026
Status	Research preview
Eligible models	Released alongside Claude Opus 4.8; works with existing models as well
Available surfaces	Claude Code CLI / Desktop / VS Code extension / headless (claude -p) / Agent SDK
Available plans	Pro, Max, Team, Enterprise
Available providers	Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry
Required version	Claude Code v2.1.154 or later
Enabling Pro / Enterprise	Pro users opt in via /config; Enterprise requires an administrator to enable it

Dynamic Workflows is enabled by default on Max and Team, opt-in for Pro through the /config menu under "Dynamic workflows," and gated behind administrator approval for Enterprise. Note the breadth of provider support: this is not limited to the direct Anthropic API. AWS Bedrock, Google Cloud Vertex AI, and Microsoft Foundry — the procurement routes Japanese enterprises tend to favor — are all supported, which keeps your procurement options open.

Relationship to Claude Opus 4.8

Claude Opus 4.8, announced the same day, looks like it was at least partly designed to bring out the best in Dynamic Workflows. Opus 4.8's standard pricing stays at $5 input / $25 output per million tokens, identical to Opus 4.7. Fast Mode, however, drops to $10 / $50 — roughly three times cheaper than the previous $30 / $150. Given that Dynamic Workflows runs many agents in parallel by nature, the Fast Mode price cut reads as a deliberate "make this affordable enough to actually use" move.

The benchmarks tell a similar story. Opus 4.8 scores 88.6% on SWE-bench Verified (up from 87.6% on 4.7), 69.2% on the more difficult SWE-bench Pro (up from 64.3%), and 74.6% on Terminal-Bench 2.1 (up from 66.1%). The gains skew toward agentic coding tasks specifically.

How It Differs from Tool Use, Managed Agents, and Subagents

This is the part most people find confusing. Since 2024, Anthropic has shipped Tool Use, the Agent SDK, subagents, Skills, and Managed Agents in rapid succession — all of which sound similar at first glance. Let's situate each one.

Subagents, Skills, and Workflows Side by Side (Anthropic's Own Framing)

The Claude Code documentation publishes a comparison table that I have adapted below.

Dimension	Subagents	Skills	Workflows
What it is	A worker Claude spawns	A set of instructions Claude follows	A script the runtime executes
Who decides what runs next	Claude (turn by turn)	Claude (following the prompt)	The script
Where intermediate results live	Claude's context window	Claude's context window	Script variables
What you can reuse	Worker definitions	Instructions	The orchestration itself
Scale	A few tasks per turn	Same as subagents	Tens to hundreds of agents per run
Behavior on interruption	The turn restarts from the beginning	The turn restarts from the beginning	Resumable within the same session

Here is the key sentence from the official documentation:

A workflow moves the plan into code. With subagents and skills, Claude is the orchestrator: it decides turn by turn what to spawn next, and every result lands in Claude's context. A workflow script holds the loop, the branching, and the intermediate results itself, so Claude's context holds only the final answer.

"The plan moves into code" is the heart of it. With subagents and Skills, Claude decides what to spawn next at every conversational turn, and every result ends up in the context window. Because the context window has a hard limit, once you spawn more than ten or so agents, the conversation history starts ballooning, and Claude begins to lose track of earlier results during later turns.

Dynamic Workflows sidesteps that design bottleneck through script variables. Intermediate results live in the script's local variables, and only the final answer returns to Claude's conversational context. The upshot is that you can run hundreds of agents while keeping the conversation context clean.

How It Compares to the Agent SDK and Managed Agents

Because the lineup is easy to confuse, here is a second table.

Type	Who runs the loop	Where it's hosted	Typical use case
Tool Use (Messages API)	The developer writes the loop	Your own infra or cloud	The baseline pattern for custom tool calls
Agent SDK	The SDK abstracts the agent loop	The developer hosts it	Embedding Claude Code-like behavior into your own application
Managed Agents	Anthropic manages the runtime	Anthropic infrastructure	Production agents that run for hours or longer
Dynamic Workflows	The runtime executes a script Claude wrote	Anthropic's sandbox	Dozens to hundreds of agents coordinating within a single session

Tool Use and the Agent SDK are foundational layers for "how to build an agent" — they assume the developer is writing the agent loop. Managed Agents is a service for "handing off a long-running agent to Anthropic in full," covering sandboxes, checkpoints, and state management on your behalf.

Dynamic Workflows is neither of those. It's a mechanism for running many subagents in parallel and in stages within a single task, and its lifetime is essentially one run inside one session. Long-running agents go to Managed Agents; complex one-shot tasks go to Dynamic Workflows. That seems to be the intended division of labor.

How It Works — Phases, State, Concurrency, and Resumption

Now let's lay out the operational picture based on the documentation.

End-to-End Flow

Drawing from the official blog, here is the typical flow in chronological order:

1. The user includes the keyword "workflow" in a prompt, or invokes a bundled command like /deep-research.
2. Claude interprets the prompt and writes a JavaScript orchestration script.
3. The CLI or Desktop client asks the user to approve the plan (subject to permission mode).
4. Once approved, the runtime executes the script in the background, in an isolated environment.
5. The script spawns subagents in phases — up to 16 concurrently, and up to 1,000 in total.
6. Each subagent investigates, edits, tests, or reviews in parallel.
7. Results accumulate in script variables, and where appropriate other agents perform adversarial review.
8. Once the answer converges, a single final report returns to Claude's conversational session.

While the run is in progress, you can open a progress view with /workflows. It shows the number of agents per phase, cumulative tokens, and elapsed time. You can also interrupt or restart individual agents from there.

State Management and Resumption

A quietly important feature is resumption. The documentation puts it this way:

The runtime tracks each agent's result as the run progresses, which is what makes a run resumable within the same session.

Because the runtime tracks each agent's result as it finishes, if you pause partway through, on resume the completed agents return their cached results and only the unfinished ones run again. The caveat is that if you fully end the Claude Code session, you can't resume — the next launch starts from scratch.

Documented Limitations

Limit	Details
User input	Interactive user input during a run is not supported. If a stage needs approval, split it into separate workflows per stage.
Direct I/O	The workflow body cannot directly access the filesystem or shell. Real I/O is delegated to subagents.
Concurrency	Maximum of 16 agents at once. On machines with fewer CPU cores, this drops further.
Total count	Maximum of 1,000 agents per run, to prevent runaway loops.
Permission mode	Subagents spawned by a workflow always run in acceptEdits mode. File edits are auto-approved.
Allowlist	Subagents inherit the user's tool allowlist. Shell commands, Web Fetch, and MCP tools not on the allowlist will still trigger approval prompts mid-run.

The permission story at the bottom is where you have to be careful. Because subagents force acceptEdits, if you haven't set up your tool allowlist properly in advance, you'll either end up with unexpected edits running unsupervised or with so many prompt interruptions that the whole "background execution" benefit evaporates.

API Surface and a Minimal Example

The story splits into the Claude Code surface and the API surface.

How to Use It from Claude Code

There are two short paths.

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

That's the bundled workflow /deep-research. It runs Web searches in parallel from multiple angles, cross-checks the sources, and returns a cited report.

The other path is to include the word "workflow" in a regular prompt:

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

Claude Code highlights the word "workflow," and Claude switches into script-writing mode rather than the usual turn-by-turn handling. The script it produces is approved by the user, then runs in the background, with the result returning to the conversation.

If you like the run, you can press s in the /workflows view to save that run's script as a command. There are two save destinations:

• .claude/workflows/ (in the project): Available to everyone who clones the repo.
• ~/.claude/workflows/ (under your home directory): Available across all your projects, invisible to others.

After saving, you can invoke /<name> to re-run the same orchestration as many times as you like.

Using It from the Agent SDK and Headless

You can also use Dynamic Workflows through claude -p (headless mode) and the Agent SDK. The documentation puts it like this:

In claude -p and the Agent SDK there is no one to prompt, so tool calls follow your configured permission rules without interactive confirmation.

In other words, you can invoke workflows from CI/CD pipelines, cron jobs, or internal backends without interactive approval. If you're already using the Agent SDK, picture inserting a "write and execute a workflow" phase inside your existing agent loop.

Disabling It at the Organization Level

If an enterprise administrator wants to turn off Dynamic Workflows entirely, there are three options:

• Toggle "Dynamic workflows" off in /config.
• Set "disableWorkflows": true in ~/.claude/settings.json.
• Set the environment variable CLAUDE_CODE_DISABLE_WORKFLOWS=1.

To apply the setting across the entire organization, add "disableWorkflows": true to Managed Settings, or flip the toggle in the Claude Code Admin dashboard. All of this is spelled out in the official documentation.

Use Cases Anthropic Itself Highlights

To stay close to the source, I'll list only the examples documented in the official blog and Claude Code documentation.

1. Codebase-Wide Bug Sweep

Anthropic calls this pattern a "codebase-wide bug sweep." It explores services and repositories in parallel and spins up an independent verification agent for each finding. For codebases too large to chase down inside a single Claude session, it produces issue reports and patch proposals in bulk.

2. Large-Scale Migrations

The example cited in the official blog is Bun (the fast JavaScript runtime) migrating from Zig to Rust. Jarred Sumner, Bun's creator, reported using Dynamic Workflows to generate around 750,000 lines of Rust and get to a state where 99.8% of the existing test suite passed — all within 11 days. Framework swaps, deprecated API replacements, and language ports — tasks that span thousands of files — are now feasible inside a single run.

3. Profile-Driven Optimization Audits

A pattern where performance bottleneck discovery and patch generation run together. Multiple subagents independently propose improvements based on profiler output, and another group of agents adversarially reviews them.

4. Security Audits

The official blog uses "security audit" as an example. The workflow searches the codebase in parallel, then spins up an independent verification agent for each candidate vulnerability. The premise is that cross-checking from multiple independent vantage points produces higher confidence than a single reviewer.

5. Deep Research

The bundled /deep-research falls into this category. It runs multi-angle Web searches, cross-references the sources, and uses "votes" on each claim to perform a survival check, returning only the claims that survive, complete with citations. Useful for internal knowledge research, competitive analysis, and industry trend research.

Enterprise Adoption — Four Axes for the Decision

From here, with Anthropic's official sources as the foundation, I'll lay out the questions that Japanese executives and DX leaders should be debating. I'm not selling Anthropic, and I'm not dismissing it — the goal is a fact-based view on where it fits and where it doesn't.

1. Security and Governance

• The fact that subagents always run in acceptEdits mode deserves attention. Running them directly against a production repository requires a properly maintained tool allowlist and execution on a dedicated branch.
• Interactive approval mid-run is not supported. For CI/CD invocations, allowlist management for MCP tools and shell commands is the operational lifeline.
• Organization-level on/off control is centrally manageable via Managed Settings or the Admin dashboard. That's reassuring for internal IT.

2. Data Sovereignty

• Direct use of the Anthropic API routes through US regions by default.
• Vertex AI, Bedrock, and Microsoft Foundry sometimes let you choose Japanese regions (though model availability needs to be verified on a case-by-case basis).
• On the Team and Enterprise plans, contractual terms specify that data is not used for model training by default.

If you're injecting Dynamic Workflows into a workflow that handles sensitive data, your choice of API endpoint directly determines where your data lives. When you bring this to the board, make sure responsibility for the "which endpoint do we route through" decision is unambiguous.

3. Operational Overhead and Cost

The documentation is honest about this in one sentence:

Dynamic workflows can consume substantially more tokens than a typical Claude Code session, so we recommend starting on a scoped task to get a feel for usage in your work.

That is, expect significantly higher token consumption than a regular Claude Code session, and start with a scoped task to calibrate. Anthropic doesn't publish a specific multiplier, but given that the agent count grows by orders of magnitude beyond a normal conversation, I'd plan on 5x to 10x token consumption for initial budgeting — that's the safer assumption.

The standard cost-control playbook combines the three mechanisms Anthropic provides:

• Prompt caching, which can cut costs by up to 90% (especially effective on shared system prompts).
• Batch processing via the Message Batches API, which cuts costs by 50%.
• Routing equivalent tasks to smaller models like claude-haiku.

A pragmatic pattern is to stage the workflow script: "heavy reasoning on Opus 4.8, verification and summarization on Haiku."

4. The Research-Preview Risk

As noted up front, Dynamic Workflows is a research preview as of May 28, 2026. The API and behavior aren't frozen, and changes may land en route to GA. Rather than slotting it directly into business-critical paths, the appropriate move is to validate it on supporting workflows first — documentation generation, code review, competitive research — before promoting it deeper.

---

Free Download: Enterprise AI Decision Framework

We have published a downloadable framework for evaluating 2026 AI agent platforms — including Dynamic Workflows — across four axes: security, cost, operational overhead, and data sovereignty.

Download the framework for free

---

Options for Japanese Enterprises — Pairing ZEROCK with Dynamic Workflows

When Japanese enterprises adopt Claude for production use, the practical choice is among three options: route Dynamic Workflows straight through the US-based Anthropic API, go through a Japanese region on Bedrock, Vertex AI, or Foundry, or keep the parts that touch internal data on a separate Japan-hosted foundation.

ZEROCK, the enterprise AI platform TIMEWELL provides, targets the third option. It runs GraphRAG-based knowledge control on AWS infrastructure in Japan and centralizes cross-functional instructions through a prompt library. Combined with a setup that calls the Claude API through a Japanese endpoint, this lets you keep highly sensitive data inside Japan while still using Dynamic Workflows' parallel processing power for research and large-document workloads — a hybrid that works well in practice.

Three concrete patterns are worth considering:

1. Knowledge in Japan, inference where it fits: Internal knowledge searches go through ZEROCK's GraphRAG; external Web research or large-scale codebase audits run on Dynamic Workflows.
2. Stage-by-stage orchestration: Within the workflow, subagents that touch sensitive data are routed to the ZEROCK API, while Web searches and general verification go straight to the Anthropic API.
3. Phased migration: Lock down core operations on ZEROCK in an on-prem-leaning configuration first, then introduce Dynamic Workflows for supporting workflows, and expand into core operations once governance has caught up.

The right answer depends on your industry and the sensitivity profile of the work. Option three is not a universal winner.

How WARP NEXT Supports "Dynamic Workflows Adoption Assessments"

Deciding whether to adopt a feature like Dynamic Workflows — and if so, which workflows to start with, and how to design governance — is a lot to ask of engineering managers and internal IT alone. It also pulls in budget allocation at the executive level, business process redesign, and risk assessment.

TIMEWELL's WARP NEXT is a monthly-engagement AI consulting service led by former enterprise DX and data strategy specialists. For 2026 enterprise AI platforms, including Dynamic Workflows, we help organize questions like:

• Which of your workflows are well-suited to Dynamic Workflows' parallelism, and which are not?
• Risk assessment and the design of internal approval processes.
• Pilot workflow selection and KPI design for measuring outcomes.
• Appropriate combinations with ZEROCK, Managed Agents, and Bedrock-based deployments.
• Translating all of the above into materials ready for executive decision-making.

Staying in "wait and see — it's still a research preview" is the kind of posture that puts you six months behind a more aggressive competitor, which is the recurring story of the AI space since 2025. Watching the space and running pilot deployments aren't a tradeoff. You need a structure that lets you do both at the same time.

Then vs. Now: A Timeline of Claude's Feature Evolution

For context, here is a timeline of the agent-related features Anthropic has shipped over the past 18 months.

Time	Feature	Position
Early 2024	Tool Use (Messages API)	The baseline pattern for custom tool calls
Mid 2024 onward	Computer Use	Screen-control automation
Early 2025	Claude Code launches	Agentic coding CLI
Mid 2025	Agent SDK released	The Claude Code capabilities, exposed as a library
Late 2025	Skills, Subagents, Managed Agents	Structuring and operating agents
May 28, 2026	Dynamic Workflows + Opus 4.8	Script-driven, large-scale parallel orchestration

The shape of the evolution: "define tool calls with Tool Use" → "abstract the loop with the Agent SDK" → "divide labor with Skills and Subagents" → "outsource the operational layer to Managed Agents" → "lift orchestration itself into a script with Dynamic Workflows." Dynamic Workflows doesn't replace any of those layers — it's a new abstraction that sits at the top of the stack. Seeing it that way makes technology selection inside the company easier to reason about.

Summary

The key points:

• Dynamic Workflows is Claude's new orchestration mechanism, released as a research preview on May 28, 2026.
• The substance is a JavaScript script. Claude writes it on the spot, and a separate runtime executes it in the background.
• The difference from subagents and Skills is that "the plan becomes code." Intermediate results live in script variables; only the final answer returns to Claude's context.
• Up to 16 agents concurrently and up to 1,000 per run. Within the same session, runs are pauseable and resumable.
• It doesn't compete with Tool Use, the Agent SDK, or Managed Agents. It coexists as the top-level abstraction.
• The official use cases include codebase-wide bug sweeps, large-scale migrations, security audits, and deep research.
• The four axes for enterprise evaluation: security, data sovereignty, operational overhead, and cost. Factor in the research-preview status.
• For Japanese enterprises, the practical answer is to combine it with a domestic foundation like ZEROCK, or to route through Bedrock, Vertex AI, or Foundry.
• WARP NEXT supports the decision process itself — how to evaluate this for your company and where to start.

With Dynamic Workflows in the lineup, agentic AI has shifted phases — from "Claude thinking inside a conversation" to "Claude writing code that divides the work." The question to debate at the board level is no longer "use it or not." It's "which workflows to start with, and at what risk threshold to introduce it."

---

If You'd Like to Organize Your AI Strategy

Deciding which of the latest enterprise AI capabilities — like Dynamic Workflows — to bring into which workflows, and on what timeline, is exactly what WARP NEXT supports through monthly engagement led by former enterprise DX and data strategy specialists. If you want to use Claude safely on Japan-based servers, ZEROCK is the candidate to evaluate.

See what WARP NEXT covers　See what ZEROCK covers　Book a 30-minute consultation

---

References

1. Anthropic. "Introducing dynamic workflows in Claude Code." May 28, 2026. https://claude.com/blog/introducing-dynamic-workflows-in-claude-code
2. Anthropic. "Orchestrate subagents at scale with dynamic workflows." Claude Code Documentation. https://code.claude.com/docs/en/workflows
3. Anthropic. "Introducing Claude Opus 4.8." May 28, 2026. https://www.anthropic.com/news/claude-opus-4-8
4. Anthropic. "Claude Opus 4.8." Product page. https://www.anthropic.com/claude/opus
5. Anthropic. "Skills explained: How Skills compares to prompts, Projects, MCP, and subagents." https://claude.com/blog/skills-explained
6. Anthropic. "Building agents with the Claude Agent SDK." Engineering Blog. https://www.anthropic.com/engineering/building-agents-with-the-claude-agent-sdk
7. Anthropic. "Create custom subagents." Claude Code Documentation. https://code.claude.com/docs/en/sub-agents
8. Anthropic. "Agent SDK overview." Claude Code Documentation. https://code.claude.com/docs/en/agent-sdk/overview
9. Anthropic. "Pricing." Claude API Documentation. https://platform.claude.com/docs/en/about-claude/pricing
10. TechCrunch. "Anthropic releases Opus 4.8 with new 'dynamic workflow' tool." May 28, 2026. https://techcrunch.com/2026/05/28/anthropic-releases-opus-4-8-with-new-dynamic-workflow-tool/
11. The New Stack. "Claude Opus 4.8 is here: effort controls, dynamic workflows, cheaper fast mode, better honesty, less deception." May 28, 2026. https://thenewstack.io/claude-opus-48-release/
12. VentureBeat. "Anthropic's Claude Opus 4.8 is here with 3X cheaper fast mode and near-Mythos level alignment." May 28, 2026. https://venturebeat.com/technology/anthropics-claude-opus-4-8-is-here-with-3x-cheaper-fast-mode-and-near-mythos-level-alignment
13. MarkTechPost. "Anthropic Ships Claude Opus 4.8 Alongside Dynamic Workflows and Cheaper Fast Mode, With Workflows Capped at 1,000 Subagents." May 28, 2026. https://www.marktechpost.com/2026/05/28/anthropic-ships-claude-opus-4-8-alongside-dynamic-workflows-and-cheaper-fast-mode-with-workflows-capped-at-1000-subagents/