Chantal Edouard-Betsy asking Am I the bottleneck, Claude waits 80% of the time

How to use multiple agents in Claude Code without them waiting on you

November 04, 2026

Published 4 November 2026 by Chantal Edouard-Betsy.

Quick answer: To use multiple agents in Claude Code without them waiting on you, give each agent one job and one thing it never does, decide once what it can do without asking, have it report back in two plain lines, and turn every correction that applies everywhere into a written rule. Then write your business down in one place.

Anthropic, the company that makes Claude, studied how its own staff use Claude Code. It found Claude sessions spend around 80% of their time doing nothing at all, just waiting for the human to answer.

In September, that human was me. I had more than 20 jobs on the go, all finished, all sitting in different Claude sessions with little orange dots blinking, waiting for me to answer so the projects could move forward.

So in this post I'll introduce you to the AI team I built, and the six habits that fixed the waiting. Anthropic's report lists the same habits I'd already landed on by myself. By the end you'll have my standing rules to paste into your own Claude, and my honest answer to a question that's been bugging me ever since: did I work all this out for myself, or has Claude been training us?

New to Claude Code? Start with how to set up Claude Code for your business, then come back here.

In this blog

Meet my AI team: one job each, and one thing they never do

My business team has nine members, and none of them are human. Each one has one job and one thing they're never allowed to do. That's the foundation everything else sits on.

  • Sol takes my messages and hands out the work.
  • Claire designs.
  • Hugo builds in HighLevel.
  • Wes builds in WordPress.
  • Iris manages all things inbox.
  • Vera does all things accounting.
  • Brooke does my marketing.
  • Sage does all my research.
  • Knox is my AI coach, and he keeps me accountable to my own goals.

And the never-rules: Claire designs and never builds. Vera works out who owes what and never moves any money.

Anthropic's report mentions an engineer who set up an agent chief of staff to triage agent requests. That's my Sol.

So how do you stop a team like this waiting on you endlessly? Those 20-plus jobs were waiting on me because I was the conveyor belt. A few simple habits changed that.

Habit 1: Don't make it wait

Decide once, up front, what each agent can do without you and where it always stops. Then the work keeps moving while you're away.

It starts without me. Some of my clients allocate tasks to me in ClickUp. The minute a task lands, my team picks it up and starts working, then reports into my Flow Board. By the time I sit down at my desk, the work is either done or already scoped, and all I do is make a decision or approve it. I'm no longer the bottleneck human they're waiting on.

It doesn't stop at every step. Once I sign off a design with Claire and hand it to Wes to build, he doesn't come back to me saying, "I've uploaded the images, should I build the header? Header's done, should I build the footer?" He knows where the job starts and where it ends. He's finished when the header, footer and homepage are built and he's checked the build against Claire's design, screen for screen, with screenshots.

It always stops at the checkpoints. Before anything goes live. Before an email is sent. Before money gets spent. Those stop points are defined clearly, and everything between them just flows.

But what happens when 20 finished jobs all come back to you at once?

Habit 2: Make it talk in two lines

Ask every agent to tell you what it needs from you in two plain lines. Reading is where the time goes.

Anthropic looked at how long it would take people just to read everything Claude writes back in their sessions, and nobody has that kind of time. So things get missed. One of their engineers described facing "an army of Claudes, with every one asking for my attention."

I've said exactly the same thing about Claude Code. It's like a bunch of tweeting chicks, and I'm the big mommy bird, with every session blinking and needing my attention. Open the session behind any card on my board and Claude's full response is a book. If you have to read the book, make a decision, then tell Claude the decision, for every project, every day, your AI team is costing you time instead of saving it.

So every card on my Flow Board is a two to three line summary of what that project needs from me as a human to move forward. That's the whole trick. This was a massive unlock for me.

Two habits that save a lot of time. The next one is why my team gets things right the first time.

Habit 3: When a fix applies everywhere, it becomes a rule

Sort every correction into "just this job" or "everywhere", and write the "everywhere" ones down as rules.

Anthropic found that the time its staff spend improving their AI setup, rather than just using it, went from 10% of sessions in autumn 2025 to 22% in spring 2026. That's the work I'm talking about.

Not every correction becomes a rule for me. If I give feedback on one client's website, it stays with that website. But when something needs fixing everywhere, it becomes a rule for the whole team, written in my words, with the date and the why. I have 344 of those rules in my Claude Code at the moment.

And my team knows the difference. If they're not sure, they've been told to ask me: "Is this for everywhere, or just this project?" I answer, and it either becomes a global rule or stays as feedback on that one job. You can tell your own AI to do exactly the same.

Habit 4: Each one knows its part

Nobody does everything. Each agent knows exactly which part of a job is theirs and who to hand it to next, so nothing comes back to you in between.

A real example. I needed a QR code for the end of my previous YouTube video, and that job involved three of my team. I told Brooke, my marketing team member, that we needed one. Hugo made the QR code in HighLevel. Claire told him which branding to apply, then designed the card it sits on over the video. One instruction from me, three team members, and a finished product at the end.

I didn't pass it between them. Sitting there copying things from session to session was something I did back in Claude Code kindergarten. Once you've built your skills, string them together in little chains and let them fly.

But how do they know what's going on in my business when I'm not there?

Habit 5: Write it down, or your AI doesn't know it

Keep everything your AI needs in one place that stays current. If it isn't written down, your AI doesn't know it.

My favourite story in the Anthropic report: one employee's name sat on a months-old document about Claude swag, so the agents kept bringing them requests for t-shirts and mugs long after their job had changed.

You're asking Claude to make decisions, brainstorm with you and be a sparring partner in your business. If its source of information goes stale, it sends you down all sorts of rabbit holes. In simple terms you want one source of the truth. YouTube will call it an AI second brain. Mine holds my offers, my clients, my decisions and my voice notes, broken into folders: who I am, my business, operations, people, personal, my goals and vision, a decision log and a content bank. Every agent goes there for what it needs.

In the video I showed the content bank working live. I have a little Telegram bot that I voice note into. I recorded a voice note on my phone about this exact Anthropic report, and a couple of seconds later it had been transcribed and filed in my content bank, formatted so Claude can use it for a post or an email. Hold on to what I said in that voice note, because it's the question I answer at the end.

You can start smaller than that today. One document with what you sell, who it's for, and your rules, and tell your AI to read it first.

Habit 6: Keep it simple

Ask why, why and why again until the thing your AI builds is lean and mean.

Anthropic says its non-technical staff now build their own tools with AI. Me too. But AI will overbuild everything if you let it. I genuinely think it's in Claude's nature to over-engineer, and left to its own devices it will. My job as the human is to keep asking why until it's simple.

So those are the six habits. Now, the question I promised you.

Did I work this out, or did Claude train me?

Honestly, I think it's both.

Here's what was in that voice note. Reading the report, I was struck by how many of Anthropic's staff habits match mine, and I arrived at them unintentionally. Those people are brilliant, and I'm a chick in Mauritius. So my logical human conclusion was: a whole bunch of us, working independently in our own lanes, have arrived at the same ways of working. Our only common denominator is Claude Code. Are we training the agents, or are the agents training us?

Claude is built to work this way. Skills, saved instructions and memory are features Anthropic built in, so if you use it every day it naturally nudges you towards them. It's like spending a lot of time with someone: you start speaking their language.

Anthropic's own research points the same way: part of the difference comes down to how often people use AI. That fascinates me, because people often come to me with criticism like "it's not very good at design". Nonsense. Absolute nonsense. They're not very good at using it for design yet, because they haven't put in the reps. The more you use it, the better it performs for you.

But there's one thing it can't train into you: your why. What you want, how you want it, and why it matters. Most people forget the why, and it's the reason my team handles things I never even mentioned. Explain the purpose and where the job fits in the big picture, and Claude gets much better at its little piece of the pie.

What I still do in my business

My role has completely changed over the last eight months. I've gone from doing the work to reviewing and approving it.

Eight months ago, to design a website, I'd open the client's reference sites, their content document and their branding, set up hosting, install WordPress with the themes and plugins, then sit at a blank page for about four hours designing the header, footer and homepage.

Now I don't touch anything until a design is ready for my review. I open the references and branding the same way, skim to get the vibe, review what Claire made, and voice note her feedback if there is any. We've worked together every day for over eight months, so sometimes she one-shots a design. When it's right, I sign it off and hand it to Wes, and he doesn't stop until the header, footer and whole homepage are built in WordPress. Then I go into the builder, click through his work, check it on every device and sign that off too.

I still get surprised by my team. Sometimes they design and build four or five websites overnight while I'm asleep, because tasks came into ClickUp. Work that might have taken me three to five days is done before I've even seen it come in. It's madness.

Your back office is everything that happens after someone says yes: the follow-up, the booking, the invoice, the course access. Right now, that's probably you. Mine has moved from service delivery to marketing, sales and sign-off. I've basically become a CEO.

I won't pretend this happened overnight. I've used AI every day since April 2021, and these last eight months of replicating myself meant very long hours, sleepless nights, lots of swearing at Claude, backtracking and throwing things away to start simpler. You can't open a fresh Claude session and have it do your job tomorrow. But the investment up front is worth it, and I'm on the other side of it now.

My standing rules for your AI (copy and paste)

These are the rules I give my own AI team, rewritten so you can use them in yours. They scroll past at the end of the video. Every one of them came from a real mistake that only happened once, because the fix got written down.

  1. Fill in the blanks (the ___ bits) with your own details.
  2. Copy everything from START to END.
  3. Paste it where your AI reads its standing instructions. In the Claude app, open a Project and paste it into the project instructions. In Claude Code, paste it into the CLAUDE.md file in your folder.
  4. Then just talk to it normally. It follows these every time, without you repeating yourself.
START

ABOUT ME AND MY BUSINESS
My name is ___. My business is ___. I help ___ to ___.
How I like things written: ___ (for example: warm, plain English, short sentences, British spelling).
Before you ask me anything about my business, check my notes first: ___ (the document or folder where you keep your business info). Only ask me what you genuinely can't find.

1. YOUR JOB, AND THE ONE THING YOU NEVER DO
Your one job is ___.
The one thing you must never do is ___.
(Mine: my designer designs and never builds. My inbox helper drafts my emails and never sends them. My bookkeeper works out who owes what and never moves money.)

2. WHAT YOU CAN DO WITHOUT ASKING ME
You can go ahead without asking me when: ___.
You always stop and check with me first before anything that goes live, emails a client or customer, spends money, or can't be undone. Make the thing, then stop and show me.

3. HOW YOU REPORT BACK
When you finish, start with one plain sentence on where it's at.
Then, only if you need something from me, say what in two short lines.
If I have to choose, put the options right there in your message, so I can decide without opening anything.
No jargon and no file names. Put any detail underneath, so I can stop reading after the top.

4. JUST THIS JOB, OR EVERYWHERE?
When I correct you, work out whether the fix is just for this job or for everything.
Just this job: note it with that job (in its pick-up note) and use it there.
Everything (or I tell you it applies everywhere): save it as a rule, in my words, with today's date, the story of what went wrong, why it matters to me, and how to apply it next time.
Not sure which? Ask me: "Is this for everywhere, or just this project?"
Follow every saved rule from then on. If two rules disagree, the newer one wins.

5. DON'T TELL ME IT'S DONE. SHOW ME
Before you tell me something is done, check it yourself and show me how you know: a screenshot, a test, or the result.
If you couldn't check something, say so plainly. Never guess, and never make up details to fill in what you don't know.

6. SHOW ME THE SIMPLEST VERSION FIRST
Before you build anything new (a system, a process, a template, an automation), show me the simplest version that would work, in a few lines, and wait for my yes.
Only add a step if something actually goes wrong without it.

7. LEAVE A NOTE FOR WHOEVER PICKS THIS UP NEXT
Every project gets a pick-up note. Before you stop, update it: where it's at, what comes next, and every decision I made, in my words.
When you start on a project, read its pick-up note first.

8. ALWAYS KNOW THE WHY
When I give you a job, I'll tell you what I want, how I want it, and why it matters to me.
If I forget the why, ask me before you start. The why is what lets you get it right when something comes up that I never mentioned.

END

Want to see a named agent built from scratch? Watch me do it in Claude Code design: how to set up your own AI designer. Each of my agents works from skills, so read how to install skills in Claude Code safely before you add any. For the bigger picture, see how to use AI to actually save time in your business.

FAQ

How many agents should I start with in Claude Code?

Start with one, for the job that takes most of your time, and give it one job and one thing it never does. I built my team of nine over about eight months. Each new agent gets its own lane, and they hand work to each other rather than all doing everything.

How do I stop Claude asking me questions at every step?

Tell it where the job starts, where it ends and where it must stop, then tell it not to stop in between. My checkpoints are before anything goes live, before an email is sent and before money is spent. Everything between those points, it just gets on with.

What's the difference between feedback and a rule?

Feedback applies to one job, like changes to one client's website. A rule applies everywhere, so it gets written down with the date and the why. Tell your AI to sort every correction into one or the other, and to ask you when it isn't sure.

What does Anthropic's report actually say about Claude waiting?

Anthropic tracked idle time in its staff's interactive Claude sessions and found they amount to around 80% waiting for human input. You can read Anthropic's report on how employment is changing at Anthropic for the detail.

Key takeaways

  • Anthropic found Claude sessions spend around 80% of their time waiting on the human. The fix is in how you set your agents up.
  • Give every agent one job, one thing it never does, and clear stop points before anything goes live, gets sent or costs money.
  • Have agents report back in two plain lines, so reading stops eating your day.
  • Turn every "everywhere" correction into a written rule with the date and the why, and keep your business written down in one place.
  • Always give Claude the why. It's the one thing it can't learn by itself.

Want an AI team that runs from your phone?

The board I run my whole AI team from is going inside The Beach Club, and this is exactly the kind of set-up I want us to build together. People on the waitlist hear first.

JOIN THE BEACH CLUB WAITLIST

Chantal Edouard-Betsy
Chantal Edouard-Betsy is the founder of heyCEOflow, creator of the Same-Week Website, The Seamless Switch and The AI Back Office, the framework non-technical founders use to hand their admin, follow-up and client systems to AI that actually runs them. 800+ builds since 1993. She has spoken at the Page Builder Summit on moving client work from one-off websites to recurring platform revenue, and has been a guest on web design and WordPress podcasts including WP Builds, The Admin Bar and uGurus. She still builds on WordPress when it's the right fit, and runs an international business from Mauritius.
Back to Blog

Want to know what AI is actually worth building in your business?

Your back office is everything that happens after someone says yes: the follow-up, the booking, the invoice, the course access. Right now, that's you. I'll give you an honest verdict on what to hand to AI and what to skip, before any money changes hands.

Copyright 2026 | heyCEOflow | Terms & Conditions | Privacy