Claude Code Ultra Code in Practice: One Person Orchestrating 100+ Agents in Parallel Development

What Are Ultra Code and Dynamic Workflow

Claude Code recently launched a major feature — Ultra Code, which, combined with Dynamic Workflow, enables a single developer to orchestrate over a hundred Agents working in parallel on large-scale tasks. Work that previously required a team two to three days can now be completed in 30 minutes.

Dynamic Workflow

The official definition: Dynamic Workflow is a JavaScript script used to coordinate sub-agents at scale.

In plain terms, traditional workflow tools (like Coze, n8n, etc.) require you to manually define every step, with agents executing along predetermined paths. Dynamic Workflow eliminates the manual definition step, letting the large model understand your intent and automatically generate an intelligent workflow. You simply describe your requirements in natural language, and it adaptively orchestrates tasks.

From a technical implementation perspective, the core innovation of Dynamic Workflow lies in upgrading traditional DAG (Directed Acyclic Graph) task orchestration to runtime-generated execution plans. Traditional workflow engines like Apache Airflow and Temporal require developers to predefine task dependencies, while Dynamic Workflow leverages the reasoning capabilities of large models to parse task structures at runtime, dynamically determining parallelism, dependencies, and error recovery strategies. Its underlying implementation as JavaScript scripts means it has full programming capabilities — loops, conditional branches, exception handling, state persistence, and more — far exceeding the linear execution model of traditional Prompt Chains.

Ultra Code: Intelligent Trigger Decision

Ultra Code is essentially a trigger decision mechanism. Workflow is the execution engine, while Ultra Code decides when to activate Workflow and when not to. When you input a complex task, Claude Code acts like a commander-in-chief, automatically determining whether multiple Agents need to work in parallel and selecting appropriate models for different subtasks.

This multi-Agent parallel scheduling is the application of distributed systems design principles in the AI domain. Similar to how the MapReduce framework splits large tasks into parallelizable smaller tasks, Ultra Code decomposes complex requirements into multiple independent or loosely-dependent subtasks, assigning them to different Agent instances for simultaneous execution. Each Agent has its own independent context window and tool-calling permissions, avoiding the attention dilution problem inherent in single long contexts (where the model's attention to information in the middle decreases as context length increases). The key challenges of this architecture include granularity control in task decomposition, consistency guarantees during result aggregation, and retry mechanisms for failed tasks.

How to Enable and Configure

Version Check

First, confirm your Claude Code version is at least 2.1.162. Use this command to check:

claude --version

If your version is too low, upgrade to the latest version first.

Enabling Ultra Code and Dynamic Workflow

In a Claude Code conversation, type:

/configure

You'll see two configuration options:

• Dynamic Workflow → Select True
• Ultra Code → Select True

Verifying Activation

Type an instruction like "I want to use Ultra Code to analyze an article." If the words "Ultra Code" appear highlighted in brown, the feature has been successfully enabled.

Practical Demo: Deep Research Workflow

Starting Deep Research

Deep Research is a built-in official workflow command in Claude Code, specifically designed for research-type tasks. It conducts multi-angle web searches around a question, scrapes materials, cross-validates sources, votes to filter conclusions, and finally returns a research report with citations.

From a methodological perspective, Deep Research implements a complete information retrieval-verification-synthesis pipeline, drawing from the systematic literature review methodology in academia. Multiple Agents handle different stages: query expansion (decomposing the original question into multiple search angles), information scraping (accessing multiple sources in parallel), fact-checking (cross-validating claim consistency across different sources), and voting mechanisms (multiple Agents independently scoring the credibility of conclusions). This design effectively mitigates the hallucination problem of single LLMs, improving output reliability through multi-source verification.

It's recommended to enable maximum permission mode to avoid frequent permission confirmation prompts:

claude --dangerously-skip-permissions

Then input:

/deep-research 给我探寻这个GitHub上的开源项目 [项目URL]，生成整体报告，教会我如何使用和修改它

Process and Results

During execution, you can use the /workflow command to check progress. In actual testing, a single Deep Research call deployed 104 agents for targeted research, completed fact-checking based on 22 sources and 59 claims, and ultimately generated a comprehensive research report.

However, the cost is significant — this single conversation consumed approximately ¥8 (about $1.10) using DeepSeek, and would start at $20 if using the Claude Opus model. Token consumption is the core cost consideration for multi-Agent systems: each Agent instance requires its own system prompt, task description, and tool-calling context, and these "fixed overheads" grow linearly with the number of Agents. Taking 104 Agents as an example, the repetition of system prompts alone could consume hundreds of thousands of tokens. Additionally, aggregating results between Agents and generating the final report requires injecting all subtask outputs into the main Agent's context window, creating additional token consumption. This explains why a single Deep Research call can cost dozens of times more than a regular conversation.

Differences Between Workflow, Sub Agents, and Skills

Many people confuse these three concepts. Here's a clear comparison:

Feature	Sub Agents	Skills	Workflow
Essence	Specialized execution units spawned by Claude	Fixed instruction templates/tool constraints	Execution plans written in code
Flow Control	Claude decides the next step	Claude progresses based on prompts	Script controls loops, branches, scheduling
Use Cases	Delegating a small number of tasks	Enhancing capabilities for certain task types	Repeatable, long-chain, recoverable tasks
Core Value	Makes Claude better at doing things	Makes Claude better at doing things	Automates the task process itself

In simple terms, Sub Agents and Skills make Claude "better at doing things," while Workflow "automates the task process itself," supporting multiple rounds of experimentation, review, strategy comparison, and quality control.

To elaborate: Sub Agents are like a manager delegating tasks to subordinates, where each sub-Agent executes specific tasks in an independent sandbox and returns results. Skills are more like equipping Claude with specific "skill cards," enhancing its performance on certain task types through predefined prompt templates and tool constraints (e.g., code review skills, documentation writing skills, etc.). Workflow is actual program code with deterministic flow control capabilities, able to handle complex conditional logic, error retries, state checkpoints, and task recovery — suitable for production-grade tasks requiring reliability guarantees.

Two Token-Saving Tips

Tip 1: Let Claude Decide Whether to Enable Workflow

When entering a project, type:

/ultracode

Claude Code will default to using Workflow mode only for complex tasks, rather than enabling it for everything. This is suitable for the learning phase, but once you're experienced, manual judgment is more reliable.

Tip 2: Mixed Model Strategy to Reduce Costs

Let cheaper models handle breadth scanning (tasks that don't require deep thinking), and let the primary model handle consolidation and logical reasoning. For example:

Ultra Code 用HiCode廉价模型铺广度，用Opus强模型收口，帮我看看这个项目的Skills有哪些改进

In testing, this command launched 9 Agents — 8 using the cheaper HiCode model to traverse skill files, and 1 using the Opus model to generate the final report — significantly reducing token consumption.

The mixed model strategy (Model Routing) is an important optimization technique in current AI engineering. Its core idea originates from cache hierarchy design in computer architecture — using low-cost resources to handle large volumes of simple requests, and high-performance resources for small numbers of complex requests. In the LLM context, different models have varying strengths in reasoning depth, context understanding, code generation, etc., with price differences of 10-50x. For example, structured tasks like file traversal and information extraction show similar accuracy between small and large models, but at one-tenth the cost. Through intelligent routing, systems can reduce total costs by 60%-80% while maintaining output quality.

Saving and Reusing Workflows

Saving a Workflow

After running a workflow, type /workflow to view the execution process. A shortcut key S (Save) will appear at the bottom. After clicking it, the system saves the workflow as a JavaScript file, with the default path being the .claude/workflow/ folder in the project root directory.

Reusing a Saved Workflow

Rename the saved file (e.g., test.js), then invoke it with natural language:

用 .claude 目录下 workflow 文件夹下的 test.js 工作流分析这个项目

Claude will ask whether to enable the workflow. After confirmation, you can reuse the previous orchestration logic. If you know JavaScript, you can directly modify the code to fine-tune the process; if not, just use natural language to ask Claude Code to modify it for you.

The saveable and reusable nature of workflows is their key advantage over regular conversations. In software engineering, this is analogous to upgrading temporary scripts into maintainable automation pipelines (CI/CD Pipelines). The saved JavaScript files contain complete task decomposition logic, Agent scheduling strategies, result aggregation rules, and error handling mechanisms, which can be reused across projects, version-controlled, and even shared within teams. This means senior developers on a team can design workflow templates for junior members to call directly, enabling standardized knowledge transfer.

When to Use It

Scenarios suitable for Ultra Code:

• Large-scale projects with tens of thousands of lines of code
• Security audits
• Large-scale code migrations
• Multi-strategy solution comparisons
• Highly repetitive complex tasks

Scenarios where it's not needed:

• Looking up a simple function
• Small requirement changes
• Low-cost, low-risk daily tasks

The core value of Ultra Code lies in giving individual developers team-level parallel collaboration capabilities, but its token consumption is also at team level. Making sound judgments about when to use it is key to truly achieving the efficiency advantage of "one person equals a whole team."

Thinking from an ROI (Return on Investment) perspective: Ultra Code is worthwhile when a task's labor cost (time × hourly rate) far exceeds the API call cost. For example, a security audit might require a senior engineer to spend two days (at ¥3,000/day = ¥6,000), while Ultra Code could complete an initial scan in 30 minutes at an API cost of ¥100-200. But for a small change that takes 5 minutes to do manually, launching a multi-Agent system would only add unnecessary complexity and wait time. The key is developing intuition for task complexity — when you find yourself needing to jump between multiple files repeatedly, compare multiple approaches, or handle large numbers of repetitive subtasks, that's when Ultra Code delivers its value.

Key Takeaways