If you are new to Claude Code, turn off dynamic workflows before they burn through your usage limit.
Last week in my workshop, this caught a few learners off guard. One prompt spun up 100+ agents to research "AI patterns to avoid in web design," and burned t...
Last week in my workshop, this caught a few learners off guard. One prompt spun up 100+ agents to research “AI patterns to avoid in web design,” and burned through their Claude usage limit about 15 minutes into the workshop. We had to stop everyone and switch it off on the spot just to keep the session moving.
Here is what happened.
On 28 May, Anthropic shipped dynamic workflows alongside Claude Opus 4.8. Instead of working through your task turn by turn, Claude can now write a script that fans out to many subagents at once. A single run can spawn up to 1,000 agents.
For the right job, this is genuinely useful. Codebase-wide audits. 500-file migrations. Research questions where every source gets cross-checked against the others. Big work that one conversation cannot hold.
For a beginner who did not ask for it, it is a footgun.
The catch: dynamic workflows are ON by default.
How to turn it off:
-
Type /config and press Enter.
-
Use the up and down arrow keys to move to the “Dynamic workflows” row.
-
Press Enter (or the spacebar) to switch it off.
-
Press Esc to close the menu.
That persists across sessions. Power users can also set “disableWorkflows”: true in ~/.claude/settings.json, or export CLAUDE_CODE_DISABLE_WORKFLOWS=1.
This is the expert-in-the-loop point in one feature. 1,000 agents do not make the answer better. They multiply whatever judgment is driving them. The expert decides when a task is actually big enough to fan out. A beginner who triggers it by accident just pays more for the same average answer.
So before you switch it on: if you run codebase audits, migrations, or deep research, learn it well and it earns its cost. If you are still learning the basics, keep it off until you know exactly when you want 1,000 agents working at once.
Docs: https://lnkd.in/gqD5HP9T
Small things like this are easy to miss when you are starting out. If you are new to Claude Code and want to learn the foundations hands-on, I run a Foundations of Claude Code workshop. The next cohorts are on 18 June (2 slots left) and 2 July.
Details: https://lnkd.in/e9r8kScw
#AI #ClaudeCode #AIAgents
Enjoyed this? Subscribe for more.
Practical insights on AI, growth, and independent learning. No spam.
More in AI Agents
What is an AI agent?
If you're still confused, you're not alone. There is no universally agreed-upon definition of what an AI agent is.
Your AI agent will get prompt injected sooner or later, because it is easier than most people thought.
Most people think prompt injection needs a carefully crafted adversarial prompt by an experienced hacker. It does not. Someone who understands how LLMs work ...
Why Your OpenClaw Agent Is One Message Away from Getting Hacked?
A stranger sent a very long, sophisticated-looking message to her agent. It was filled with detailed research instructions about finance news, complete with ...
OpenClaw vs Claude vs Hermes vs NanoClaw
If you have used Claude Code heavily, you saw it immediately: OpenClaw is a chat wrapper on top of CLI agents like Claude Code and Codex. So I dismissed it a...
Most OpenClaw workshops skip the fundamentals and teach you 1-click setup on AI or cloud providers...
I decided to go for a different path. In my workshop, I guide learners to set up with the original experience: install it on your laptop, but with security i...
Isn't RAG just a glorified search engine?
This was my thought when I first read about it. I still think it is.
What is an AI agent?
If you're still confused, you're not alone. There is no universally agreed-upon definition of what an AI agent is.
Why Your OpenClaw Agent Is One Message Away from Getting Hacked?
A stranger sent a very long, sophisticated-looking message to her agent. It was filled with detailed research instructions about finance news, complete with ...
Most OpenClaw workshops skip the fundamentals and teach you 1-click setup on AI or cloud providers...
I decided to go for a different path. In my workshop, I guide learners to set up with the original experience: install it on your laptop, but with security i...
Your AI agent will get prompt injected sooner or later, because it is easier than most people thought.
Most people think prompt injection needs a carefully crafted adversarial prompt by an experienced hacker. It does not. Someone who understands how LLMs work ...
OpenClaw vs Claude vs Hermes vs NanoClaw
If you have used Claude Code heavily, you saw it immediately: OpenClaw is a chat wrapper on top of CLI agents like Claude Code and Codex. So I dismissed it a...
Isn't RAG just a glorified search engine?
This was my thought when I first read about it. I still think it is.