Claude Code’s Dynamic Workflows Eliminate Manual Task Splitting – A Hands‑On Test

Dynamic Workflows in Claude Code

On 28 May 2026 Claude Code released the dynamic‑workflow feature (research preview). It lets the model automatically split a task, run dozens to hundreds of sub‑agents in parallel, and cross‑validate their results.

Key capabilities

Automatic task decomposition – Claude analyses the request and creates subtasks without user‑defined workflow.

Large‑scale parallel execution – 10‑100+ agents can run simultaneously, giving a 2‑10× speed improvement (e.g., a 2‑3 h article written in 5 min, a 100‑file code review reduced from 1‑2 days to ~30 min).

Cross‑validation – independent agents verify each other's output, improving quality and reducing errors.

Iterative convergence – the system repeats the process until answers converge.

Checkpoint‑resume – execution can be paused and resumed later.

Enabling dynamic workflows

Two methods:

Enable the ultracode setting (sets effort level to xhigh) via shortcut Ctrl+Shift+E (or Cmd+Shift+E on macOS) or by issuing the command /effort ultracode. After enabling, describe the task; Claude decides whether a dynamic workflow is needed.

Directly request a workflow, e.g.

创建一个动态工作流，帮我写一篇关于 Claude Code 动态工作流的公众号文章

. Claude then plans, creates an orchestration script, runs the agents, cross‑validates, and returns the final answer.

Step‑by‑step example (article generation)

我想写一篇关于 Claude Code 动态工作流的公众号文章

Claude’s automatically generated workflow performed:

Analyse task complexity.

Decide to use a dynamic workflow.

Split into multiple subtasks (information gathering, drafting, polishing).

Run three agents in parallel for data collection.

Cross‑validate the collected information.

Iteratively converge on a final draft.

The whole process completed in under five minutes, compared with an estimated two‑to‑three hours for manual authoring.

Comparison with traditional serial execution

Large‑scale tasks (hundreds of files, multi‑day migrations) benefit from parallelism.

Simple single‑session tasks or tasks with strong inter‑dependencies are better handled without dynamic workflows.

Token‑cost‑sensitive or latency‑critical scenarios may avoid dynamic workflows because the startup adds overhead.

Benchmark examples reported:

Write a public‑account article: 2‑3 h → 5 min.

Code review of 100 files: 1‑2 days → ~30 min.

Technology‑stack migration: several weeks → several days.

Practical guidelines

Start with a small task to measure token usage, e.g.

# Small test
创建一个工作流，帮我写一个简单的 Python 脚本

# Larger task after measurement
创建一个工作流，帮我重构整个用户模块

Monitor progress with the /workflows command, which shows running workflows, stage progress, and agent status.

Enable auto mode ( /auto) so Claude automatically confirms intermediate steps, reducing manual interaction.

Design tasks for parallelism – e.g., assign one agent per file for refactoring, or multiple agents to test different hypotheses during debugging.

Control token cost :

Simple tasks – single session.

Medium tasks – Agent Team (3‑5 agents).

Large tasks – Dynamic workflow (10‑100+ agents).

Insert manual confirmation at critical nodes, e.g.

创建一个工作流，帮我重构用户模块，在合并代码前让我确认

Applicable scenarios

Massive code refactoring (hundreds of files).

Cross‑file bug hunting.

Technology‑stack migration.

Code audit and cleanup.

End‑to‑end feature development.

Content creation such as articles.

Quick start

# Enable ultracode
/effort ultracode

# Describe task
我想重构 xxx 模块的 XXX 类

# Monitor workflow
/workflows

Note: dynamic‑workflow mode consumes more tokens than a regular session.