Introducing Dynamic Workflows in Claude Code: Quarter-Long Work in Days

Update — July 27, 2026: Claude Code can now use Sakana AI’s Fugu model orchestra as its backend. See the Fugu interface setup, pricing, and compatibility limits.

TL;DR: Dynamic workflows in Claude Code let Claude orchestrate tens to hundreds of parallel subagents to tackle complex engineering tasks end-to-end. Work that normally takes quarters now finishes in days—bug hunts across entire services, migrations touching hundreds of files, and plans stress-tested from every angle.

What Are Dynamic Workflows?

Anthropic launched dynamic workflows in Claude Code today (May 28, 2026), introducing a paradigm shift in how AI handles complex, large-scale engineering tasks.

Some problems are too big for one pass by a single agent, especially in complex, legacy codebases:

• Bug hunts across an entire service
• Migrations that touch hundreds of files
• Plans you want stress-tested from every angle before committing

Dynamic workflows handle all of these end-to-end.

How Dynamic Workflows Work

When a workflow kicks off, Claude:

1. Plans dynamically based on your prompt
2. Breaks it into subtasks with clear objectives
3. Fans the work out across subagents running in parallel
4. Checks results before they're folded in
5. Iterates until answers converge

Agents address the problem from independent angles, other agents try to refute what they found, and the run keeps iterating until the answers converge—which is how a workflow reaches results a single pass can't.

Key Capabilities

Parallel Execution: Dynamic workflows can run dozens to hundreds of subagents simultaneously, each working on independent subtasks.

Adversarial Checking: Independent agents work to break the results before you see them, ensuring high-quality outputs when the cost of a wrong answer is high.

Long-Running Sessions: Built for work that can extend into hours and days, with progress saved continuously so interrupted jobs pick up where they left off.

Coordinated Results: Because the coordination happens outside the conversation, the plan stays on track no matter how big the task gets.

Real-World Use Cases

Early access users and teams inside Anthropic have been using dynamic workflows for:

Codebase-Wide Audits

• Bug hunts: Claude searches the entire service or repo in parallel, then runs independent verification on every finding so the report surfaces real issues
• Security audits: Hardening passes for auth checks, input validation, and unsafe patterns across an entire codebase
• Performance optimization: Profiler-guided optimization audits with parallel analysis and verification

Large-Scale Migrations

• Framework swaps: Migrating from one framework to another across thousands of files
• API deprecations: Updating deprecated API calls throughout the codebase
• Language ports: Converting entire codebases from one language to another while maintaining test coverage

Critical Work Verification

When the cost of a wrong answer is high, a workflow gives Claude:

• Independent attempts at the problem
• Adversarial agents working to break the result
• Multiple verification passes before you see the output

Case Study: Rewriting Bun from Zig to Rust

Jarred Sumner used dynamic workflows to port Bun from Zig to Rust—a monumental task that showcases what workflows can unlock at scale:

The Numbers:

• ~750,000 lines of Rust generated
• 99.8% test suite passing after port
• 11 days from first commit to merge
• Hundreds of agents working in parallel

The Process:

1. Lifetime mapping workflow: Mapped the right Rust lifetime for every struct field in the Zig codebase
2. Port workflow: Wrote every .rs file as a behavior-identical port of its .zig counterpart, with two reviewers on each file
3. Fix loop workflow: Drove the build and test suite until both ran clean
4. Optimization workflow: Addressed unnecessary data copies overnight and opened PRs for final review

While not yet in production, all of this was handled by dynamic workflows. Jarred will be writing about this more in the future.

How to Enable Dynamic Workflows

Dynamic workflows are available today in research preview for:

• Claude Code CLI
• Claude Code Desktop
• VS Code extension (Max, Team, and Enterprise plans if admin enabled)
• Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry

Option 1: Direct Request

Ask Claude to create a dynamic workflow:

"Create a workflow to audit all authentication checks in our codebase"

Option 2: Ultracode Mode

Turn on the ultracode setting (accessible through the effort menu):

• Sets effort level to xhigh
• Lets Claude decide automatically when to use a workflow
• Best for ongoing complex work

For the best experience, turn on auto mode when using dynamic workflows.

Important Considerations

Token Consumption

Dynamic workflows can consume substantially more tokens than a typical Claude Code session. Anthropic recommends:

• Start on a scoped task to get a feel for usage
• The first time a workflow triggers, Claude Code shows what's about to run and asks you to confirm
• Organization admins can optionally disable workflows through managed settings

Availability by Plan

When to Use Dynamic Workflows

Perfect For:

• Migrations touching 100+ files
• Security/performance audits across entire repos
• Bug hunts in complex, legacy codebases
• Critical decisions needing multiple verification passes
• Language ports requiring test suite validation

Not Ideal For:

• Single-file changes
• Simple refactoring
• Quick bug fixes
• Exploratory coding

Getting Started

1. Ensure you're on a compatible plan (Max, Team, or Enterprise)
2. Enable auto mode for the best experience
3. Either:
   • Ask Claude to create a workflow directly, or
   • Enable the ultracode setting through the effort menu
4. Start with a scoped task to understand token usage
5. Review the plan when Claude asks for confirmation

Read the official documentation to learn more.

What This Means for Engineering Teams

Dynamic workflows represent a fundamental shift in how AI can handle engineering work:

From: Single-threaded AI assistance on isolated tasks To: Orchestrated AI workforce tackling quarter-long projects in days

Impact:

• Migrations that were previously deferred due to scope become feasible
• Security audits can be comprehensive rather than sample-based
• Legacy modernization accelerates from months to weeks
• Critical architectural decisions get stress-tested before implementation

The coordination happens outside the conversation, so the plan stays on track no matter how big the task gets.

Try Dynamic Workflows Today

curl -fsSL https://claude.ai/install.sh | bash

Or read the documentation to learn more.

Update — July 16, 2026: Jarred Sumner's Bun Zig→Rust port — ~50 workflows, ~64 peak agents, 11 days — is the largest public dynamic-workflow case study. Full breakdown: Bun rewrite Fireship coverage.

Update — July 20, 2026: Claude Code ships Bun 1.4 Rust runtime — Simon Willison's strings verification guide for the stealth June rollout.

Update — September 17, 2026: Ultracode is now documented as a Claude Code setting rather than an effort level, with its own settings key, a size guideline, and hard runtime caps (16 concurrent agents, 1,000 per run). See Ultracode in Claude Code: what it actually does for the current behaviour, including why the Large workflow cost warning is suppressed while ultracode is on.

What changed since launch

This post documented the May 28, 2026 research preview. Several specifics have since moved, and the current documentation contradicts parts of the original announcement:

Availability widened. Dynamic workflows are no longer gated to Max, Team, and Enterprise. They are available on all paid plans, plus Anthropic API access, Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry. On Pro you turn them on from the Dynamic workflows row in /config rather than getting them on by default.

Hard runtime caps are now published. The runtime allows up to 16 concurrent agents (fewer on machines with fewer CPUs, including CPU-limited containers), a maximum of 4,096 items in a single parallel() or pipeline() call, and 1,000 agents total per run. Scripts cannot load modules: any script containing import() fails before the run starts.

A size guideline now shapes how big Claude builds a workflow. workflowSizeGuideline takes small (fewer than 5 agents), medium (fewer than 10, the default), large (fewer than 50), or unrestricted. Pro plans default to small on v2.1.271 and later. It is advice sent to the model, not an enforced cap.

Cost warnings arrived. Claude Code flags a run scheduling more than 25 agents, or projected past 1.5 million tokens, with a Large workflow warning. The warning is advisory and does not pause the run, and it is suppressed entirely in sessions with ultracode on.

Usage-limit behaviour improved. On v2.1.271 and later, agents that hit a claude.ai usage limit pause and wait for the reset rather than failing, up to twice per run, in interactive subscription sessions.

The keyword got locked down. Since v2.1.210 the ultracode keyword triggers a workflow only from prompts a human typed. It no longer fires from -p prompts, scheduled tasks, webhook payloads, or pull request comments relayed into the conversation.

For the current behaviour in full, see Ultracode in Claude Code: what it actually does.

Related Reading

• Ultracode: xhigh effort plus automatic orchestration — the setting that drives these workflows
• Bun Zig→Rust AI rewrite — dynamic workflows at million-line scale
• Claude Code Loops Official Guide (July 2026) — proactive loops compose dynamic workflows with /schedule and /goal
• Loop Engineering with Claude Code
• Agency Agents: 144+ AI Specialists to Transform Your Workflow
• Code Modernization with Claude Code
• Claude for Enterprise Teams

This post covers the research preview that began May 28, 2026; see "What changed since launch" above for current availability, caps, and cost controls.

Original note: Token consumption can be significantly higher than typical sessions—start with scoped tasks to understand usage patterns for your work.*