Dynamic Workflows in Claude Code: Hundreds of Parallel Subagents

If you've watched Claude Code chew through a hard task for an hour and wished you could fan it out to twenty agents in parallel, today's announcement is the one you wanted. Anthropic just introduced Dynamic Workflows in Claude Code, and the framing is direct: this is the feature for problems "too big for one pass by a single agent."

The post on claude.com describes Claude Code workflows that "dynamically write orchestration scripts that run tens to hundreds of parallel subagents in a single session, checking its work before anything reaches you." Claude plans the fan-out, runs the subagents, has other subagents try to refute their findings, iterates until the answers converge, and hands you the result.

This guide covers what Dynamic Workflows actually does, the problem it was built for, how to turn it on, the real examples Anthropic published (including a 750,000-line Zig-to-Rust port and Klarna's dead-code discovery), how it compares to Auto Mode, Agent Teams, and Scheduled Tasks, and when to reach for it. Dynamic Workflows ships alongside the Claude Opus 4.8 release.

The Problem Dynamic Workflows Was Built For

The official announcement is direct about the pain. "Some problems are too big for one pass by a single agent, especially in complex, legacy codebases: a bug hunt across an entire service, a migration that touches hundreds of files, a plan you want stress-tested from every angle before you commit to it."

If you've ever asked Claude to audit a service for security issues, you know the shape of the problem. A single agent reads files sequentially, builds context, and hits a ceiling. Either it loses the thread of what it already checked, or it compacts the conversation and forgets findings from earlier files. By the time it produces a final report, you're not confident it actually looked at everything.

The workaround until today was to spawn subagents manually. You'd write a slash command that delegates per module, then aggregate the results yourself. That puts the orchestration burden on you: you have to know in advance how many subagents to spawn, how to split the work, and how to merge findings. For a one-time bug hunt across a service you've never touched, you don't have that knowledge yet.

Dynamic Workflows fills the gap. You describe the task. Claude decides how to split it, how many subagents to spawn, and how to verify results before showing them to you.

What Dynamic Workflows Actually Does

The mechanics, per the announcement: Claude "plans dynamically based on your prompt, breaks it into subtasks, and fans the work out across subagents running in parallel. Results are checked before they're folded in."

That last part is the differentiator. Dynamic Workflows isn't just parallel execution. Anthropic describes a verification loop where "agents address the problem from independent angles, other agents try to refute what they found, and the run keeps iterating until the answers converge."

Three things matter here:

Parallel fan-out. A workflow can run "tens to hundreds of parallel subagents in a single session." This is the scale jump. Most Claude Code multi-agent patterns top out at four or five subagents because that's what a human can hold in working memory. Claude doesn't have that limit when it's the one orchestrating.

Adversarial verification. Subagents don't just report findings. Other subagents are tasked with refuting them. If a subagent claims a function has a race condition, another subagent's job is to disprove it. Only claims that survive the refutation pass reach you.

Convergence-driven iteration. The workflow keeps running until the answers stop changing. This is different from a fixed-step pipeline. The number of subagents and the number of iterations are decided in real time based on what the work actually requires.

The result is that a single Dynamic Workflows session can do work that previously required either a much longer single-agent run or a hand-orchestrated multi-agent setup.

Availability and Status

This is one of the broader release windows Anthropic has shipped for a Claude Code feature. VS Code, Bedrock, Vertex, and Foundry on launch day signals Anthropic is treating Dynamic Workflows as core infrastructure, not a CLI-only experiment.

The "research preview" label means token usage, behavior, and the activation interface may evolve. The announcement notes that "the first time a workflow triggers, Claude Code shows what's about to run and asks you to confirm." That confirmation step is a safety valve, not a permanent UX choice.

How to Use Dynamic Workflows

Activation has two paths.

Ask Claude Directly

The simplest path is to tell Claude what you want and use the trigger phrase.

claude
> Create a workflow that audits the auth module for race conditions across all of our service endpoints.

Claude reads the request, decides this fits the Dynamic Workflows pattern, plans the fan-out, and shows you what's about to run. You confirm, and the subagents kick off.

For users on Max or Team plans, this works out of the box. For Enterprise users, your admin needs to enable workflows first.

Enable the ultracode Effort Setting

The second path is to enable workflows as a default behavior. Open the effort menu in Claude Code and set the level to ultracode. This sets effort to xhigh and lets Claude decide on its own when a request warrants a workflow versus a normal session.

With ultracode on, you don't have to use the "create a workflow" phrasing. Claude evaluates each request and reaches for Dynamic Workflows when the task scope justifies it.

Pair It With Auto Mode

Anthropic's explicit recommendation: "Turn on auto mode when using dynamic workflows." If you haven't enabled Auto Mode yet, read our guide on Auto Mode for the full setup. The short version: Auto Mode is the permission mode that lets Claude execute actions without prompting you on every tool call, with an AI classifier blocking actions that look dangerous.

Why does Anthropic want them paired? Because a workflow that runs hundreds of subagents and stops on every permission prompt isn't actually parallel. You'd spend the whole run clicking approve. Auto Mode keeps the subagents moving while the classifier handles the safety calls.

Confirm Before First Run

The first time a workflow triggers in your session, Claude Code surfaces what's about to run and waits for confirmation. This is the safety floor against accidentally kicking off a 200-subagent session when you only wanted a single answer.

After the first confirmation, subsequent workflows in the same session execute without re-prompting, unless the scope changes substantially.

Real Examples From the Announcement

Anthropic published several concrete use cases. These are not hypothetical.

The Bun Zig-to-Rust port. Anthropic points to Bun's recent rewrite as the most extreme example: "750,000 lines of Rust, and eleven days from first commit to merge," with "99.8% of the existing test suite passing." Anthropic's framing: "Work you'd normally plan in quarters now finishes in days." A systems-level port at that scale only makes sense if Claude is orchestrating hundreds of parallel subagents with verification loops catching regressions before they merge. For the broader patterns Claude Code uses on codebases at this scale beyond Dynamic Workflows, see the large codebase playbook.

Klarna's dead-code discovery. Alessio Vallero, Senior Engineering Manager at Klarna: "Dynamic workflows have been especially valuable for discovery and review tasks across large codebases. We've seen strong results using it to identify dead code and surface cleanup opportunities that traditional static analysis missed, helping our engineers move faster on maintenance and refactoring work." The key phrase is "that traditional static analysis missed." Static analyzers operate on syntax. Dynamic Workflows can reason about whether code is actually reachable in production paths.

CyberAgent on positioning. Ken Takao, Lead Systems Engineer at CyberAgent: "Dynamic workflows fill the gap between firing off a single subagent and building out a full agent team. Plan to implementation just flows, so we can trust longer runs without losing visibility." Single subagents are too small for service-level work. Hand-built Agent Teams require upfront orchestration design. Dynamic Workflows lives between them.

Other categories Anthropic flags as good fits: codebase-wide bug hunts, security audits, large migrations spanning thousands of files, and "critical work you need checked twice" where wrong answers are costly.

Dynamic Workflows vs Other Claude Code Autonomy Features

Claude Code has shipped a series of autonomy features over the past year. They're easy to confuse, so here's how Dynamic Workflows fits among them.

Auto Mode is orthogonal. It controls how each tool call gets approved. A Dynamic Workflows session runs inside whatever permission mode you've set. Anthropic recommends Auto Mode because default mode would surface permission prompts for every subagent's actions and stall the parallelism.

Agent Teams require upfront design. With Agent Teams, you define the roster ahead of time, Frontend Specialist, Backend Engineer, Quality Engineer, each with a role description. That's right when the work decomposes cleanly into named domains. Dynamic Workflows is right when you don't know the decomposition yet.

Scheduled Tasks operate on a different axis. Scheduled Tasks and Routines decide cadence. Dynamic Workflows is about depth. You can combine them: a Routine that fires a Dynamic Workflows audit every Sunday night is a reasonable architecture.

The "tens to hundreds" number matters. A hand-built sub-agent system rarely scales past five concurrent agents. Dynamic Workflows operates at a different order of magnitude because Claude does the orchestration accounting.

When to Use Dynamic Workflows (and When Not To)

The token-cost section of the announcement is explicit: "Dynamic workflows can consume substantially more tokens than a typical Claude Code session." Anthropic's recommendation: "Start on a scoped task to get a feel for usage." Treat the first few runs as experiments to calibrate your budget before pointing it at a full codebase. Anthropic also raised the Claude Code rate limits alongside the Opus 4.8 launch to absorb the heavier workflow consumption, so if you have hit the previous ceiling before, the new headroom matters; see our usage optimization guide for patterns that keep workflow runs inside their token envelope.

Use Dynamic Workflows When

• You're doing a codebase-wide audit (security, dead code, performance) where coverage matters more than speed.
• You're planning a migration that touches hundreds or thousands of files and you want every change verified before it merges.
• You're stress-testing a plan from multiple angles before committing (architecture reviews, refactor proposals).
• The work needs the adversarial verification loop. If a wrong answer is costly, the refutation pass is worth the tokens.
• You're working on a problem in unfamiliar territory and don't yet know how to decompose it into named roles.
• You want the Bun-style "quarters become days" timeline for a port or rewrite.

Skip Dynamic Workflows When

• The task is small enough for a single agent to handle in one pass. Token cost outweighs the benefit.
• You already know the role decomposition. If you can name "Frontend, Backend, Quality" in advance, Agent Teams gives you tighter control.
• You're cost-sensitive on this session. Workflows consume substantially more tokens, and there's no premium tier separation, you're spending against your normal usage allowance.
• The work needs strict determinism. Dynamic Workflows decides the fan-out at runtime. If you need the same plan every time, hand-orchestrate it.
• You're on Pro. Workflows is a Max, Team, or Enterprise feature. Pro users will hit the gate.

The Orchestration Shift

The deeper change Dynamic Workflows represents is that the orchestration step is now a model decision, not a developer decision. Previously, you wrote the orchestration logic in slash commands, prompt templates, or sub-agent files. Now you write the request, and Claude writes the orchestration.

This matters for how you think about agentic engineering. The job moves up a level. You're no longer writing the orchestration. You're writing the success criteria, the constraints, and the trust boundaries. The orchestration follows.

A reasonable first-run pattern to calibrate your token budget:

claude --permission-mode auto
> Create a workflow that audits just the auth/middleware/ folder for race conditions. Report findings without making changes.

Narrow enough to bound the cost. Once you see how the workflow decomposes the task and how many subagents it spawns, you have a calibration point for larger work.

Frequently Asked Questions

Does Dynamic Workflows require Opus 4.8?

The announcement does not gate Dynamic Workflows on a specific model. It ships across Claude Code's existing model lineup. That said, Dynamic Workflows launches alongside the Claude Opus 4.8 release, and the kind of high-stakes work that justifies workflow-level fan-out (security audits, large migrations, port projects) is exactly the work Opus 4.8 was tuned for. If you're using workflows for the use cases Anthropic highlights, you're likely already on Opus.

Can I use Dynamic Workflows on the Pro plan?

No. Workflows is a Max, Team, and Enterprise feature. Pro is not in the supported plans list. If you're on Pro and want workflow-style orchestration, your options are to upgrade or to hand-build sub-agent systems with slash commands.

What's the difference between Dynamic Workflows and "create a workflow" in earlier Claude Code versions?

This is the first time Claude Code has shipped first-class Dynamic Workflows as a feature. Earlier sessions that used phrases like "create a workflow" were just informal language for sub-agent orchestration. The new feature gives that phrase concrete meaning: Claude plans, fans out across tens to hundreds of subagents, runs adversarial verification, and iterates until convergence.

Can my admin block Dynamic Workflows for our org?

Yes. Organization admins can disable workflows for their team. For Enterprise customers, workflows ship disabled by default, the admin has to opt in.

Does Auto Mode block subagents inside a Dynamic Workflow?

Auto Mode's classifier evaluates each tool call inside the workflow the same way it evaluates calls in any other session. The classifier doesn't know "this came from a workflow," and it shouldn't. The safety floor is per-action, not per-session-type. If a subagent in a workflow tries to run a destructive command on production infrastructure, the classifier blocks it the same way it would in a normal session.

How does Dynamic Workflows compare to running my own multi-agent system?

For most teams, Dynamic Workflows is the better starting point. You skip the orchestration design work and get adversarial verification for free. Custom multi-agent systems still make sense when you need deterministic behavior, audit logs of exactly which agents ran when, or the ability to hot-swap individual agent implementations. For exploratory work, the workflow wins on time-to-result.

What's Next

Dynamic Workflows is the most significant change to Claude Code's autonomy model since Auto Mode went GA. Combined with Claude Opus 4.8, the practical capability ceiling for what a single session can accomplish just moved meaningfully up.

For day-one usage, the playbook is straightforward:

• Turn on Auto Mode if you haven't already.
• Run your first workflow on a scoped task to calibrate token usage.
• Pair with Agent Teams when you have a known role decomposition, and use workflows when you don't.
• Layer workflows under a Routine or Scheduled Task when the work is recurring (nightly audits, weekly migration progress reports).

The Bun port is the headline example. The Klarna dead-code discovery is the one most teams will recognize. Both prove the same thing: when you let Claude orchestrate the subagents, work that was unrealistic last week becomes a Tuesday afternoon session.

The orchestration step used to be the developer's job. As of today, it's the model's.

If you're building a Claude Code setup that pairs Dynamic Workflows with permission rules, hooks, and an agent framework already tuned for orchestration handoffs, the ClaudeFast Code Kit ships with the patterns above preconfigured.