Claude Code Dynamic Workflows¶

What It Is¶

On May 28, 2026, Anthropic announced dynamic workflows as a research preview in Claude Code. The core idea: Claude writes a JavaScript orchestration script on the fly, then a separate runtime executes that script, spinning up dozens to hundreds of subagents in parallel to tackle complex engineering tasks.

Trigger: Use the word "workflow" in a prompt, or enable /effort ultracode for automatic workflow detection.

How It Works¶

Claude receives a natural-language request
Instead of working turn-by-turn, it generates a JavaScript orchestration script
A runtime executes the script in the background
The script spawns coordinated subagents (up to 1,000 total, 16 concurrent)
Intermediate results live in script variables, not in Claude's context window
Final output is a consolidated report

Key activation paths:
- Explicit: include "workflow" in your prompt
- Automatic: /effort ultracode (Claude decides when workflows are warranted)

Key Technical Details¶

Scale: Up to 1,000 total agents per execution, 16 running concurrently
Concurrency cap: Tuned for average local machine resources
Repeatability: Scripts can be saved to .claude/workflows/ or ~/.claude/workflows/ and re-executed via slash command
Resumability: Interrupted workflows resume from where agents left off — cached results persist
Model: Each agent uses the current session's model unless the script routes to a different one
Permissions: Subagents always operate in acceptEdits mode; the script itself cannot access filesystem or shell directly
No mid-execution user input: Only permission prompts can pause a workflow

Why It Matters (vs. Classic Subagents)¶

Aspect	Classic Subagents	Dynamic Workflows
Plan location	Claude's context window	JavaScript script variables
Context consumption	Every intermediate result eats tokens	Intermediate state lives outside conversation
Scale	Limited by context window	Up to 1,000 agents
Repeatability	Manual re-run	Save and replay scripts
Resumability	Start over	Resume from cached state

Real-World Validation¶

Jarred Sumner rewrote Bun from Zig to Rust using dynamic workflows — generating ~750,000 lines of code in 11 days while keeping 99.8% of tests green. That's the flagship demo: a quarter-scale project compressed into days.

Practical Applications¶

Repository-wide audits: Check every endpoint for missing auth, scan for vulnerabilities across thousands of files
Large-scale refactoring: Automated migrations across entire codebases
Deep research: The bundled /deep-research command fans out web searches from multiple angles, cross-checks sources
Bug hunting: Distributed search across entire services
Code generation at scale: Parallel file generation with verification

Caveats¶

Token cost: Each agent pays its own context overhead — can burn substantially more tokens than standard sessions
Model matters: Switching to Opus 4.8 for a 500-agent audit changes the bill by an order of magnitude
Ultracode is session-scoped: Resets on every new session; drop back to /effort high after heavy tasks
Research preview: Behavior may change before general availability

Getting Started¶

/model opus-4.8
/effort ultracode
"Run a workflow to check every endpoint under src/routes/ that is missing authentication"

Or use the bundled /deep-research command for zero-config workflow exploration.

Bibliography¶

@ClaudeDevs (May 28, 2026). Claude Code dynamic workflows [Tweet/analysis]. X/Twitter.
Anthropic (May 28, 2026). Introducing dynamic workflows in Claude Code [Blog post]. https://www.anthropic.com
Anthropic (May 28, 2026). Claude Opus 4.8 [Announcement]. https://www.anthropic.com

sub-agent-parallelism — Parallel subagent fleets (Anthropic scales to 1,000)
subagent-spawning — Subagent spawning mechanism
workflows — The workflows concept
workflows/pipeline-with-gates — Pipeline model with quality gates
multi-agent-pipeline — Multi-agent orchestration

Source: @ClaudeDevs tweet (May 28, 2026), Anthropic blog post, and third-party coverage.