Claude Code’s Dynamic Workflows: Subagent Orchestration at Scale
I spent the morning digging into Claude Code's new dynamic workflows, and they're the most interesting orchestration feature Anthropic has shipped since subagents themselves. The pitch is simple: instead of Claude deciding turn-by-turn what to spawn next, you let Claude
I spent the morning digging into Claude Code's new dynamic workflows, and they're the most interesting orchestration feature Anthropic has shipped since subagents themselves. The pitch is simple: instead of Claude deciding turn-by-turn what to spawn next, you let Claude write a JavaScript script that orchestrates dozens of subagents in the background while your session stays responsive.
Why workflows beat manual subagent delegation
I've used subagents in Claude Code for months, but the model has a ceiling. Claude is the orchestrator — it decides what to spawn next, and every intermediate result lands in its context window. That works for a few parallel tasks, but falls apart at scale. A 500-file migration or a codebase-wide audit isn't a conversation; it's a program.
Dynamic workflows move the plan into code. The script holds the loop, the branching, and the intermediate results. Claude's context window only sees the final answer. More importantly, the orchestration itself becomes repeatable — you can save the script and rerun it, version it, or hand it to a teammate.
Here's how the three modes compare:
| Feature | Subagents | Skills | Workflows |
|---|---|---|---|
| What it is | A worker Claude spawns | Instructions Claude follows | A script the runtime executes |
| Who decides what runs next | Claude, turn by turn | Claude, following the prompt | The script |
| Where intermediate results live | Claude's context window | Claude's context window | Script variables |
| What's repeatable | The worker definition | The instructions | The orchestration itself |
| Scale | A few delegated tasks per turn | Same as subagents | Dozens to hundreds of agents per run |
| Interruption | Restarts the turn | Restarts the turn | Resumable in the same session |
The built-in workflow you can try today
The fastest way to see this in action is /deep-research, the bundled workflow that fans out web searches across several angles, fetches sources, cross-checks them, and synthesizes a cited report. Run /workflows to watch progress — you see each phase with agent counts, token totals, and elapsed time. The session stays free while the agents work.
Two ways to trigger a workflow
Ask for it explicitly. Include the word "workflow" anywhere in your prompt:
Run a workflow to audit every API endpoint under src/routes/ for missing auth checksClaude Code highlights the word and writes the script instead of working turn-by-turn. If the run works, save it as a command with s in the /workflows view.
Let Claude decide with ultracode. Set /effort ultracode and Claude plans workflows automatically for every substantive task. A single request can chain several workflows: one to understand the code, one to make the change, one to verify it. It burns more tokens and takes longer, but the quality is noticeably higher for complex tasks. I drop back to /effort high for routine work.
How a workflow runs under the hood
The script is JavaScript. Claude writes it for the task you describe, and a runtime executes it in the background. Each run gets its own state, and you can pause and resume. The current limits are reasonable for real work:
| Limit | Value |
|---|---|
| Max agents per run | 100 |
| Max tool calls per agent | 25 |
| Max run duration | 50 minutes |
Cost is tracked per run and visible in the progress view. You can stop individual agents, restart them, or pause the entire workflow.
What I'm using this for
My immediate use case is a codebase audit I run monthly: check for deprecated patterns, missing tests, and stale dependencies. Before workflows, I broke this into chunks and fed them to Claude one by one. Now I describe the audit, Claude writes a script that spawns agents for each module, and I get a consolidated report with every claim cross-checked. The orchestration is in the script, not in my prompts.
The feature is in research preview and requires Claude Code v2.1.154 or later. It's available on all paid plans, plus Anthropic API access, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. On Pro, turn it on in /config under the Dynamic workflows row.