How I Built an AI Chief of Staff to Run My Agency

I have five AI agents running my design agency right now. Not as a side project. Not as an experiment. As the actual operating system of my business.

Let me back up.

Eight tools, zero overview

I'm Louie Valkhof. I'm 29, based in Arnhem, Netherlands. I've been running Oase Creative for six years -- a branding agency focused on e-commerce. We've done 200+ projects. 176 five-star reviews on Trustpilot. Built brands from first product to seven-figure revenue.

And by early 2026, I was drowning.

Not in work. In context.

Eight tools. Slack for team communication. ClickUp for project management. Google Drive for assets. Email for client communication. WhatsApp for quick questions. LinkedIn for content. Analytics dashboards. A CRM that nobody actually kept updated.

Every morning started the same way: open laptop, check email, check WhatsApp, check project statuses, check if anyone responded on LinkedIn, check if the blog post went out, check if the freelancer delivered. By the time I'd checked everything, two hours were gone. I hadn't done any actual work. I'd just... managed inputs.

I started Oase because I wanted to build brands. Instead, I was building spreadsheets.

Meet Milo

His name is Milo. He runs on a Mac Mini in my apartment, 24/7. He's not a chatbot. He's not a wrapper around ChatGPT. He's a multi-agent system built on Claude that operates my entire agency backend.

Here's the architecture:

Dew handles research. Every morning at 06:00, before I'm awake, Dew scans RSS feeds, industry news, competitor activity, and market trends. By the time I open my eyes, there's a research brief waiting.

Bloom writes content. Blog posts, LinkedIn drafts, email campaigns. Not random AI slop -- Bloom knows my voice, my opinions, my style. Every piece gets written in my actual tone, using real data from my business.

Reed monitors communication. Email comes in, Reed classifies it. Lead? Flag it. Spam? Archive it. Client question? Draft a response. Complaint? Stop everything and alert me immediately.

Tide publishes. When I approve a blog post, Tide handles the rest -- formatting, SEO checks, pushing to the website, posting on LinkedIn, scheduling tweets.

Coral handles visuals. Carousel designs, blog cover images, infographics. The visual layer that makes content actually look like it came from a design agency.

Five agents. One system. Zero tab-switching at 7 AM.

A day in Milo's life

Here's what actually happens on a typical Tuesday.

05:45 -- Milo restarts with a clean context. Reads its own handover notes from the previous session. Picks up where it left off.

06:00 -- Dew runs the morning scan. Checks RSS feeds, scans industry blogs, looks at what's happening in e-commerce and SaaS.

07:00 -- My briefing lands. Not in Slack. Not as a notification. A structured summary in my email and on Telegram: what happened overnight, what's scheduled today, any leads that came in, content that's ready for review, and 2-3 networking suggestions with actual LinkedIn profile links.

I read this while walking on my treadmill with a coffee. No screens to manage. Just a briefing to absorb.

08:30 -- If it's a posting day, Tide publishes the pre-approved LinkedIn post. First comment gets added with the link (never in the body -- kills reach).

Throughout the day -- Reed watches the inbox. If a new lead comes in via WhatsApp, I get a Telegram message within minutes. Not "you have a new message." Instead: "New inquiry from [name], asking about Amazon listing design. Sent Calendly link. Want me to follow up?"

Every 30 minutes -- Milo checks for new triggers, processes pending tasks, makes sure nothing falls through the cracks.

This isn't theoretical. This is what's been running since early April 2026.

First numbers

First LinkedIn post after Milo took over content: 3,267 impressions. For a profile with ~1,300 followers, that's a 250% reach ratio. Not because AI wrote better copy. Because the system ensured consistency -- right topic, right time, right format, every time.

Eight blog posts live in the first two weeks. All 3,000+ words. All SEO-optimized. All in Dutch and English. Before Milo, I published maybe one blog per month when I could find the time.

Daily X/Twitter posts about building Oase Software, our SaaS platform for design agencies. Consistent for the first time ever.

The morning routine went from two hours of screen management to 15 minutes of reading a briefing while walking.

But here's what I won't tell you: it was smooth from day one.

Where it broke

The first version of the briefing was garbage. Generic summaries with no actionable data. Networking suggestions without LinkedIn links -- just names floating in space. I had to rebuild the briefing format three times.

Blog cover images didn't generate properly for the first week. The agent would write the post, skip the image, and call it done. I had to add explicit validation: if the image step was skipped, the entire output gets rejected.

The content calendar had duplicate entries because one agent would add ideas without checking what was already there. Deduplication logic had to be built in after I noticed the same topic suggested four times in one week.

Email classification had false positives. Reed flagged a legitimate client email as spam because the subject line matched a phishing pattern. That one stung.

Every single one of these failures taught me the same lesson: agents don't fail at the task. They fail at the process. The writing is fine. The scheduling is fine. The handoffs between agents -- that's where things break.

Four lessons after two weeks

AI writes the first draft. Humans make it real.

Every blog post, every LinkedIn draft, every email response -- Bloom writes the first version. I edit. Sometimes I change three words. Sometimes I rewrite half of it. But starting from a solid draft instead of a blank page? That changed everything.

The key is that Bloom knows my voice. Not because I fed it "write in a casual tone." Because it reads my actual WhatsApp messages, my real opinions, my specific positions on things like AI in design. The voice file is 460 lines of rules, examples, and anti-patterns. That specificity is what separates useful AI content from slop.

Agent-ready is a competitive advantage.

Most agencies will work with AI agents within 2-3 years. We're doing it now. That's not a flex -- it's a head start. When your competitors are still figuring out how to use ChatGPT for blog posts, you have a system that handles research, writing, publishing, monitoring, and visual content in a coordinated pipeline.

The system matters more than the model.

Everyone debates which AI model is best. That's the wrong question. The right question is: what's the workflow? How do agents hand off to each other? What happens when one fails? How does context persist across sessions?

Milo crashes. Sessions end. Context gets lost. The system handles this with handover files, a SQLite database that logs every decision and event, and a loop script that automatically restarts everything. The model is a component. The system is the product.

Quality over volume. Always.

I could have Milo publish five blog posts a week. Ten LinkedIn posts. Twenty tweets. Instead: 2 blogs per week, 2 LinkedIn posts, daily tweets. Every piece earns its place. 3-4 pieces of content per week, each one something that could only come from someone who actually runs an e-commerce branding agency.

The temptation with AI is to produce more. The discipline is to produce better.

The bigger picture

I'm not just running an agency. I'm building software.

Oase Software (getoase.com) is a platform for design agencies -- the tool I wish existed when I was juggling eight different apps. My co-founder Lam and I are building it while I run the agency. The goal: 1,000 users in 2026.

That dual perspective -- agency operator AND software builder -- is what makes Milo possible. I'm not building AI tools in theory. I'm building them because I need them. Every day. For real work. With real clients paying real money for real results.

And the two systems talk to each other. Inside Oase Software lives Dune -- our built-in AI that sits on top of real project data. Select a client, pick a channel, and Dune generates marketing content from actual briefs, deliverables, and feedback. Not templates. Real context.

Milo takes that output further. Dew researches the market. Bloom rewrites Dune's drafts in my voice. Tide publishes them. The content pipeline starts inside the software I built and ends on the platforms my audience reads. One data source, one pipeline, zero copy-pasting between tools.

That's the loop: agency work feeds the software, software feeds the content, content feeds the agency. And both systems get better every week because I'm using them for real work, not demos.

What's next

Milo is getting smarter. Not in the AI sense -- in the operational sense. Better handoffs between agents. Tighter validation on outputs. More context preserved across sessions.

The next agents on the roadmap: an outreach agent that handles LinkedIn networking autonomously. A CFO agent that tracks revenue and flags financial trends. Reddit monitoring for brand mentions.

But the core principle won't change: AI handles the system. Humans handle the judgment.

I still approve every client email. I still review every blog post. I still make every strategic decision. Milo doesn't replace me. Milo replaces the two hours of tab-switching that was eating my mornings alive.

If you're running an agency and spending more time managing tools than doing actual work, you don't need a better tool. You need a system that manages them for you.

That's what I built. And I'm just getting started.