CloudCodeTree LogoCloudCodeTree
AI NewsTutorialsAbout
CloudCodeTree Logo
CloudCodeTree
  • AI News
  • Tutorials
  • About
← Back to AI News
One Claude Code Prompt Fans Out to 100 Parallel Agents: Dynamic Workflows for Bulk Tasks

One Claude Code Prompt Fans Out to 100 Parallel Agents: Dynamic Workflows for Bulk Tasks

Chris Harper

3 min read

Aug 5, 2026 · 04:09 UTC

AI
Workflow
Claude Code
Agents

Claude Code's Dynamic Workflows run a JavaScript script that fans out to tens or hundreds of subagents in the background — you watch, pause, and resume from /workflows without touching your active conversation.

Most engineers hit Claude Code's scale ceiling when a task spans the whole codebase: audit every file for a security pattern, migrate 300 API call sites to a new client, generate tests for every module. A single conversation can't hold that. Dynamic Workflows solve it.

How it works

Tell Claude Code what you want at scale — it writes a workflow script, you approve, and a dedicated runtime executes it outside your conversation. Each agent is an independent subprocess with its own tools and context; you can fan out 10 or 100 in parallel, pipeline them through stages, and resume from a checkpoint if the run is interrupted.

# In your Claude Code session:
"Audit every TypeScript file in src/ for unsafe any-casts,
then open a draft PR for each file with more than 3"

Claude writes a workflow script roughly like:

export const meta = {
  name: "fix-any-casts",
  description: "Audit and fix unsafe any casts across src/"
}
const files = await agent("List all TypeScript files in src/", { schema: FILE_SCHEMA })
const results = await pipeline(
  files.paths,
  f => agent(`Count unsafe any casts in ${f}`, { schema: AUDIT_SCHEMA }),
  (r, f) => r.count > 3
    ? agent(`Fix any casts in ${f} and open a draft PR`)
    : null
)

Agents that return null (file had ≤3 casts) are simply skipped — no wasted turns.

Interactive controls

Open /workflows (or press W in Claude Code) to see a live progress tree. Press P to pause, P again to resume. Press X to skip a stuck agent, or retry one you want to re-run. All completed agents cache their results — resuming a paused run picks up exactly where it stopped without re-running finished work.

When to use vs. a normal session

Use a dynamic workflow when: (a) the task is too large for one context window, (b) the work is naturally parallel (same operation across many files), or (c) you want the long-running work isolated from your active conversation. For quick, interactive back-and-forth, stay in a normal session.

Requirements: Claude Code v2.1.154 or later, any paid plan.

Sources: Dynamic Workflows — Claude Code Docs · Claude Code Workflows: Deterministic Multi-Agent Orchestration — alexop.dev · Claude Code Dynamic Workflows Complete Guide — stacknotice.com