Claude Code has just rolled out a major update — Workflow, a feature that transforms Claude Code from a solo coder into an orchestrator of agent teams. This innovation might be Anthropic's next big leap following MCP and Skill, and here's how it works in practice.
What Makes Claude Code Workflow Different?
When you think of workflow tools, Dify or Coze might come to mind. But Claude Code's Workflow is unique:
- Describe a task, and Claude automatically writes a JavaScript orchestration script.
- It breaks tasks into stages, assigns each stage to different agents, and coordinates dozens to hundreds of subagents in the background to deliver results.
Unlike the temporary Agent Teams of the past, Workflow solidifies agent coordination into code. It becomes auditable, reusable, and version-controllable — no more relying on the model's on-the-spot performance.
Deep Dive into the Built-in Deep-Research Workflow
Let's take the official deep-research workflow as an example. It consists of five stages:
1. Scope: Multidimensional Disassembly
Break down your question into search terms from 5 different angles to ensure comprehensive coverage.
2. Search: Parallel Search
5 agents search simultaneously, each fetching 4-6 results for their assigned angle — no waiting for each other.
3. Fetch-Extract: Deduplication and Extraction
After deduplication, up to 15 webpages are crawled. Each page yields 2-5 verifiable concrete conclusions.
4. Verify: Cross-Validation
For each conclusion, 3 skeptical agents independently look for counterevidence. Conclusions with 2 or more objections are rejected.
5. Synthesize: Merge and Synthesize
Merge verified conclusions, sort them by confidence, and compile a final report with source citations.
Practical Demo: Researching Claude Code Workflow
To test this, we used deep-research to investigate Claude Code's latest Workflow feature and write an article. Here's what happened:
- Time: 25 minutes
- Agents Involved: 104
- Token Usage: 2.8 million
- Claims Verified: 25
- Output: A complete article covering Workflow's core positioning, differences from Agent Teams, runtime limits, and the built-in
deep-researchworkflow.
Budget Control: Avoid Token Overspending
Without budget limits, Workflow could become a "token shredder." To prevent this, you can define a budget directly in the task:
workflow 100kClaude Code will optimize agent count, stage depth, and convergence method based on this budget.
How to Use Workflow
There are two ways to activate Workflow:
1. Trigger with a Prompt Word
Mention "workflow" in your prompt. The word will turn colorful, and Claude will automatically generate a workflow:
请 workflow 帮我深度调研 Claude Code 的 Workflow 功能，写一篇文章2. Command Activation
Use /effort ultracode to let Claude generate a workflow based on your task:
/effort ultracodeTo check the running workflow and agent count, use:
/workflowsThe Significance of Workflow
Claude Code's Workflow isn't just a new feature — it's a signal that AI programming is shifting from individual capability to organizational capability.
- In the past, we asked: Can a model write code?
- Later, we asked: Can an agent complete a task independently?
- Now, we ask: Can a system organize multiple agents to complete complex tasks verifiably within a budget?
This shift opens up new possibilities for tackling large-scale, multi-faceted challenges with AI.
常见问题
How is Workflow different from regular multi-agent setups in Claude Code?
The key difference: Workflow is code-driven, not prompt-driven. Traditional multi-agent setups rely on the model deciding at runtime which agents to spawn. Workflow generates a JavaScript script that defines the exact orchestration logic — stages, agent assignments, parallelization, verification gates. This script is auditable (you can read exactly what it will do), reusable (run the same workflow on different inputs), and version-controllable (commit it to Git). It's the difference between improvising a team on the spot and having a written playbook.
What's a reasonable budget for a deep-research workflow?
Based on the demo: a single deep-research task with 104 agents consumed 2.8M tokens. For quick overviews, budget 500k-1M tokens. For thorough research with adversarial verification (the default), budget 2-3M tokens. Start with a smaller budget and increase if you need deeper coverage. The budget cap is a hard limit — the workflow will optimize within it rather than exceeding it.
Can I write my own custom workflows?
Yes. Workflow scripts are plain JavaScript with a defined API (agent(), parallel(), pipeline(), phase(), log()). You can write them by hand or have Claude Code generate them from a natural language description. Custom workflows can be saved to .claude/workflows/ and reused across sessions. The API supports patterns like pipeline (sequential stages), parallel (concurrent agents), and loop-until-dry (iterative discovery).