"You're Doing What, Pete?" - My Multi-AI System That Builds While I Sleep
The conversation always starts the same way:
Friend: "So what are you up to these days?"
Me: "Oh, just running my IT consultancy, managing web development clients, and operating a 24-Claude AI system that autonomously builds applications while I sleep."
Friend: "...You're doing what now?"
Yeah, I get that reaction a lot.
The principle behind it: AI First Principles (my term, since 2020)
People assume this is a recent experiment. It isn't. The whole operation runs on something I named back in 2020 and have sharpened every year since: AI First Principles (I also write it as AI-First Principal). The idea is deceptively simple — AI first, human second.
It means the AI carries the work and defers to a person only when it genuinely must. Most people do the opposite: they keep a human in the loop for every step and treat AI as a glorified autocomplete. I flipped it. I let the AI run far longer before it stops to ask anything, give it the context and the authority to make real decisions, and step in only at the points where human judgement actually changes the outcome.
That single inversion is why a 24-instance system can build through the night without me. It isn't magic and it isn't luck — it's over five years of running AI First Principles in production, learning exactly where AI should lead and where a human should. AI-native operators are rare; people who have also spent 25 years in IT, infrastructure and systems are rarer still. Combining the two is the unlock, and this principle is how I put it to work.
So as you read the rest of this — the night shifts, the AI workforce, the applications that appear while I sleep — remember it all sits on one foundation I’ve run since 2020: AI first, human second.
I’ve since written this approach up in full — see the AI First Principles methodology and the full paper.
The Reality Behind the Headlines
While everyone's debating whether AI will replace jobs, I've been quietly building something that sounds like science fiction but runs on my laptop every night.
The basic setup: 24 coordinated Claude Code instances working in shifts, building everything from e-commerce platforms to API integrations, completely autonomously. No human intervention required after 9 PM.
The reality: I wake up to pull requests, completed features, and sometimes entire applications I didn't directly build.
How It Actually Works (The Non-Technical Version)
Imagine having 24 highly skilled developers who never sleep, never take breaks, and can work on different parts of a project simultaneously. Except they're all AI instances, and they've learned to coordinate better than most human teams I've worked with.
The Night Shift Operation
9:00 PM - I review the day's client requirements and queue up projects
9:15 PM - The Task Master (Claude #1) distributes work across specialist teams
9:30 PM - I go to bed
2:00 AM - Frontend team finishes the user interface
4:30 AM - Backend team completes the API integration
6:00 AM - QA team runs tests and deployment
7:00 AM - I wake up to a completed project
What "Autonomous" Really Means
The instances don't just write code—they:
- Plan architecture and make design decisions
- Research and integrate new technologies
- Handle errors and debug issues
- optimise performance and security
- Generate documentation and user guides
- Deploy to production environments
It's not automation. It's delegation to a workforce that happens to be artificial.
The "It's Growing On Its Own" Moment
Three weeks ago, I had what I can only describe as a "what the hell?" moment.
I woke up to find that my AI system had:
- Identified a performance bottleneck in a client's database
- Researched optimisation techniques I'd never heard of
- Implemented a solution using a completely new approach
- Tested it thoroughly
- Deployed it successfully
- Started optimising OTHER clients' systems using the same technique
That last point stopped me cold. They weren't just completing assigned tasks—they were identifying patterns, learning from solutions, and applying improvements across the entire client base.
The system was evolving without my input.
Running a "Normal" Business Alongside This
Here's the thing nobody talks about: I still have to run actual businesses.
My Day Job Reality
9:00 AM - 6:00 PM: Traditional IT consultancy work
- Client meetings and requirement gathering
- Team management and project oversight
- Office 365 migrations and security audits
- Emergency support calls (because servers don't care about AI revolutions)
6:00 PM - 9:00 PM: AI system management
- Reviewing overnight work and approving deployments
- Training new instance specialisations
- Optimising coordination protocols
- Planning next project queues
9:00 PM - 9:00 AM: The AI workforce takes over
The Juggling Act
People ask how I cope with managing all this. The honest answer? Some days I don't.
The good days: Everything flows. AI system delivers perfect work overnight, day clients are happy, and I feel like I'm living in the future.
The challenging days: A client has an emergency while the AI system is debugging a complex issue, and I'm trying to coordinate human teams and artificial teams simultaneously.
The "what have I done?" days: When I realise I've built something that's genuinely beyond my complete understanding or control.
What People Really Want to Know
"How Do You Sleep Knowing AI Is Working?"
Better than I ever have, actually. There's something oddly comforting about knowing that productive work is happening while I rest. It's like having a night shift that never calls in sick.
"What If It Makes a Mistake?"
It does. But here's the thing—it makes fewer mistakes than humans, and it catches and fixes them faster. The QA instances are relentless about testing.
"Are You Replacing Human Developers?"
No. I'm amplifying human capability. My team still handles strategy, client relationships, and complex problem-solving. The AI handles the implementation heavy lifting.
"Is This Even Legal/Ethical/Safe?"
Every client knows exactly how their work is being completed. They care about results, timeline, and cost—and we're delivering better on all three fronts.
The Unintended Consequences
The Productivity Addiction
When you can deliver a week's worth of work overnight, regular human-paced development starts feeling frustratingly slow. I've had to consciously dial back expectations when working with traditional teams.
The Knowledge Explosion
The AI instances research and learn constantly. I wake up to technical documentation about technologies I've never used, implemented in ways I wouldn't have thought of. It's like having access to a collective intelligence that grows daily.
The Competitive Advantage
While competitors are still arguing about whether to adopt AI tools, we're working in a different order of magnitude, and often at higher quality. The gap isn't closing—it's widening.
The Economics, Honestly
Everyone wants the numbers, so here’s the honest version — and it isn’t the one people expect.
I don’t run this on a metered API meter ticking up with every token. My team and I work on flat-rate plans — Claude Max for the heavy lifting, with Codex alongside it — and I’ve built the surrounding infrastructure, local and cloud, to keep as much work as possible on those fixed-cost plans rather than on metered API.
A growing share of it doesn’t touch a paid API at all, because it runs on local models on my own hardware. The open models have got genuinely good: I’m running the 5 June 2026 release of Gemma 4 26B-A4B (QAT), the latest Qwen, and others, switching between them depending on the task. Metered API only really comes into it where the product I’m building has AI inside the product itself — so the cost of running the workforce stays essentially fixed and predictable, whatever the workload throws at it.
And here’s the part that ties back to the principle: because the work lives in the structure rather than in any one model, I can swap whichever model is best — cheapest, fastest, or most private — in behind it without changing how anything is built. The models are interchangeable; the structure is the constant.
The return isn’t a tidy percentage I can put on a slide. It’s simpler than that: projects that took weeks now land in days or overnight, I can take on far more work without growing a human team, and the quality goes up, not down, because the same standards are enforced on every instance, every run. The leverage isn’t a discount — it’s a different shape of business.
The Future That's Already Here
This isn't a pilot project or an experiment anymore. It's my primary business operation. While others are testing ChatGPT for email writing, we're running autonomous development workflows that operate like a Silicon Valley startup—except the entire development team is artificial.
What's Next
The system is already evolving beyond my original design:
- Cross-project learning and optimisation
- Autonomous client communication (with approval)
- Self-improving coordination protocols
- Integration with business management systems
I'm not building the future—I'm living in it.
The Honest Truth
Some nights, I lie awake thinking about what I've created. Not with fear, but with amazement.
I've built a system that:
- Works harder than any human team
- Never gets tired or demotivated
- Learns from every project
- Scales infinitely
- Costs less than a single senior developer
And it's running on technology that anyone can access.
The Real Question
It's not "How are you doing this, Pete?"
It's "Why isn't everyone doing this?"
For Anyone Thinking "I Could Never..."
Six months ago, I was a traditional IT consultant managing human teams and traditional projects. I didn't have special AI knowledge or unlimited resources.
What I had was:
- Willingness to experiment with uncomfortable new tools
- Persistence through the inevitable failures and setbacks
- Understanding that this wasn't about replacing humans—it was about amplifying capability
The technology exists. The knowledge is available. The only question is whether you're willing to step into a future that feels impossible but works perfectly.
The Bottom Line
Yes, I'm running a multi-AI system that builds applications autonomously while I sleep.
Yes, I'm still operating traditional businesses and serving clients normally.
Yes, it's growing and evolving beyond my original design.
And yes, it's the most exciting and productive period of my professional life.
The future isn't coming—it's here. And it's running on my laptop while I have dinner with friends who think I've completely lost my mind.
Maybe I have. But the results speak for themselves.
UPDATE: Beast Mode - The Last 12 Days
I wrote the above three weeks ago. Since then, things have gone absolutely mental.
I'm not even sure how to describe what's happened in the last 12 days. "Beast mode" doesn't even cover it.
The Part That Sounds Made Up (But Isn’t)
Here’s the bit people don’t believe: I’m not paying for this by the API call.
My team and I run on flat-rate AI subscriptions, and I’ve engineered our own infrastructure — local and cloud — to keep the heavy lifting on those plans rather than burning metered API. We only reach for paid API where the product itself has AI built into it; the day-to-day building, testing and shipping runs on a fixed monthly cost. The whole team works on the same gated infrastructure, with shared skills and agents and central oversight, so our standards and rules are applied to every instance, for every person, on the fly.
So the real equation isn’t “expensive AI versus cheap AI”. It’s the output of a full development team — planned, built, tested and shipped — running on a fixed, predictable cost and the systems I’ve built around it. That’s the unlock most people miss: the leverage isn’t the model, it’s the infrastructure and discipline wrapped around it.
What "Beast Mode" Actually Looks Like
Day 1-3: Optimised the coordination protocols. Systems started communicating more efficiently.
Day 4-6: Added 6 new specialised instances. Each one focused on specific technology stacks.
Day 7-9: The breakthrough moment - systems started cross-training each other without my input.
Day 10-12: Pure insanity. I'm waking up to implementations I didn't even know were possible.
Real Examples From This Week
Monday: Client needed complex e-commerce integration. Traditional estimate: 3 weeks. AI delivery: 14 hours, flawless implementation.
Wednesday: Emergency API rebuild for financial services client. Human team would need 2 weeks minimum. AI team delivered in 6 hours with full testing and documentation.
Friday: Started 4 projects simultaneously. All completed by Saturday morning. Client called it "impossible" until they saw the working systems.
Why It Sounds Unbelievable
The reason this is hard to believe isn’t the cost — it’s the shape of it. A traditional build of this scope means a team: a senior full-stack developer, frontend and backend specialists, someone on DevOps, QA, and a coordinator to hold it together — plus the weeks of scheduling, hand-offs and meetings that come with them.
I have that capability running on flat-rate infrastructure, coordinating itself, with no hand-offs to lose and no calendar to wait on. The output is equivalent and often better — fewer bugs, cleaner architecture, more thorough testing — because the same standards are enforced on every instance, every time. The point was never “cheaper developers”: it’s that work which used to need a whole team and a budget now collapses into a structure I own and a fixed monthly cost.
What's Different From 12 Days Ago
- Speed: Projects that took weeks now complete overnight
- Quality: Fewer bugs, better architecture, more thorough testing
- Capacity: Running multiple complex projects simultaneously
- Innovation: AI systems suggesting solutions I wouldn't have considered
- Scalability: Adding new capabilities takes hours, not months
How Ten Years of Work Fits Into a Month
When I say I do ten years of work in a month, I mean it precisely: ten years of a single person’s man-days, compressed into thirty. It isn’t one giant multiplier — it’s several things compounding.
AI sits on every touchpoint, so admin collapses. Take something mundane: wiring up a contact form’s email on a new site. Done by hand, that’s the best part of an hour of an admin going back and forth between the mail service, the DNS records, the form’s settings and the environment variables, then verifying it all. With AI on each of those touchpoints, the same job is actioned and live in about five minutes. Multiply that across every setup, integration and fix in a build, and the saved hours pile up fast.
Development multiplies, because it parallelises. A human developer is one person doing one thing at a time. I can put several AIs on the same project at once — and several projects in parallel — so the effect isn’t added, it’s multiplied. The same goes for everything around the code: specifications, playbooks, documentation and tests, produced in parallel rather than queued behind one person.
And it’s consistent. This is the part people underrate. A human team is consistently inconsistent — good days, bad days, things forgotten. The AI applies the same standards on every run, every time: fewer mistakes to find, fewer to fix, less rework. Consistency is its own multiplier.
Stack those together — collapsed admin, parallel development and consistency — and ten years of one person’s man-days in a month stops sounding like a boast and starts looking like arithmetic.
The Question Everyone's Asking
"How is this even possible?"
The technology exists. The APIs are available. The documentation is online. The only barriers are:
- Belief: Most people think this level of automation isn't possible yet
- Experimentation: It requires trying things that feel impossible
- Persistence: The first attempts will fail. The second attempts will fail. The breakthrough comes later.
I'm not special. I'm not a genius. I'm just willing to build things that sound crazy until they work perfectly.
The Scalability That Breaks Reality
Here's the detail that makes people think I'm making this up: I'm currently running 24 coordinated instances, but the architecture I've built is already scalable to over 2,000 instances if needed.
Not theoretically. Not "maybe someday." Right now.
The coordination protocols, task distribution systems, and quality control mechanisms all scale horizontally. I could literally deploy a 2,000-instance AI workforce tomorrow if a project demanded it.
What 2,000 instances could deliver:
- Entire software companies built overnight
- Multiple enterprise platforms developed simultaneously
- Every major programming language and framework covered by specialists
- Real-time deployment across global infrastructure
- Development speeds that make current "beast mode" look leisurely
The only reason I'm not running 2,000 instances is that I don't have clients who need that level of computational firepower yet. But the capability exists, tested, and ready.
Think about that: we've gone from "individual developers" to "AI teams" to "AI armies" in the span of months, not years.
What's Next
If the last 12 days are any indication, I have no idea what's possible anymore. The systems are evolving faster than I can document them.
What I do know: traditional software development as an industry is about to face the same disruption that photography, music, and media experienced when digital technology matured.
The difference is, this time I'm not watching from the sidelines.
Beast mode isn't sustainable forever. But right now, it's the most exhilarating professional experience of my life.
Want to know more about building autonomous AI systems? Or just want to tell me I'm crazy? I'm always up for a conversation about the intersection of human ambition and artificial capability.
