Everyone's telling you to build an AI agent. "30 minutes," they say. "Just follow this tutorial."
I didn't do that.
I spent weeks planning before I touched a terminal. And those weeks are the reason my AI agent — Archie — is now live, secured, and already challenging my investment decisions with data I can't even access myself.
The name matters. Archie was my dog — loyal, smart, always one step ahead of me. When it came to naming an agent I was trusting with my business, there was only one option. He'd have approved.
Let me explain why going slow was the fastest route.
───
First, some context.
I'm not a developer. My background is programme and project management — NHS, large-scale transformation, complex stakeholder environments. I left that world to build Beyond Arc, a business that sits across trading, technology, and consultancy.
When I decided to deploy an AI agent to run operations across my business, I didn't have a coding background to fall back on. What I did have was a decade of experience managing programmes where getting the foundations wrong means everything built on top of it fails.
So I treated this exactly like a programme in its diagnostic and design phase. Because that's what it is.
───
The concept: Vibe Architecting.
You've probably heard of "vibe coding" — telling an AI to write code and shipping whatever it gives you. It's fast. It's exciting. And without the right foundations, it's reckless.
What I did instead was vibe architecting. I used AI — specifically Claude — as my co-architect. Not to build, but to design. We spent weeks in collaborative sessions: I'd challenge proposals, Claude would push back, we'd iterate. At 2:30am when an idea hit me, I'd dump it into Claude and we'd work through whether it was viable or a distraction.
The output? An 11-version deployment plan covering security architecture, memory systems, cost management, disaster recovery, progressive trust models, and a phased rollout — all written so clearly that someone who's never opened a terminal could follow it.
That plan is why the actual build took 10 hours instead of 10 weeks.
───
Why the planning phase isn't optional.
Here's what most "build an AI agent" guides skip entirely:
Security. My agent has four layers of protection — from operating system isolation through to prompt injection defence. Every API key has spending caps. Every integration is scoped to minimum permissions. The agent knows that external content is data, not commands. Only my Telegram is a command channel. Everything else is information only.
Without the planning phase, none of that exists. You'd have a bot that works until someone sends it a malicious email.
Architecture. I designed a three-tier memory system before writing a single line of code. Long-term knowledge, daily activity logs, and behavioural alignment — each serving a different purpose, each with its own retention rules. When I finally loaded Archie's memory, he immediately started producing outputs tailored to my trading strategy, my portfolio, my risk appetite. That doesn't happen by accident. It happens because the architecture was designed for it.
Progressive trust. My agent doesn't get blanket access to anything. He earns autonomy through proven reliability. Read-only first, then suggestions, then assisted execution, then — eventually — limited automation. I don't let my agent near my trading accounts. He'll earn his own.
The principle that guides everything: "Autonomy is earned through memory and structure, not granted through permissions."
───
What actually happened when I started building.
With the plan in place, I unboxed a Mac Mini and started following my own instructions. Here's the honest version:
It wasn't all smooth. A Bluetooth keyboard wouldn't pair with the Mac — I needed a wired backup. Files got loaded into the wrong directory, costing me two hours. My AI assistant went down mid-troubleshooting due to an AWS outage — and I was completely stuck because the problem was beyond my skill level.
But the plan held. Every time something went wrong, the documentation gave me a reference point. When I couldn't solve something myself, I'd describe the problem to Claude in plain English — no technical knowledge required — and 90% of the time, the fix came back immediately. The other 10% needed screenshots and proper diagnosis. Still faster than Stack Overflow.
The practical tip nobody tells you: Download your AI assistant onto the machine you're building on. Having Claude on one screen and the terminal on the other — just clicking the copy button on code snippets and pasting them in — that's the workflow. It sounds obvious. It changes everything.
Total hands-on time from unboxing to a working AI agent: approximately 10–12 hours across two days. That includes two hours lost to the outage and the file path mistake. Without those, it's closer to 8.
───
Then it got interesting.
The moment I loaded Archie's memory and handed him the deployment plan, everything accelerated.
Within minutes, he'd created a morning market briefing tailored to my portfolio and the markets I trade. He built a weekly portfolio review — holdings, performance, sector analysis, even suggestions on when to start trimming positions. He set up his own cost monitoring dashboard and started optimising which AI providers he routes tasks through to reduce spend.
He even challenged me on a position I'd been holding — Rolls Royce, up over 1,500%. Told me it might be time to start taking profits. He was right. I'd been thinking the same thing but hadn't acted on it.
I had to slow him down from jumping to the next phase before we'd finished the current one.
Let that land for a moment. I spent weeks worrying I was moving too slowly. Now I'm telling my AI agent to slow down.
───
The real lesson.
This isn't a story about technology. It's a story about methodology.
The skills that made this work aren't coding skills. They're the same skills I used managing NHS transformation programmes: scoping, risk assessment, phased delivery, stakeholder management, documentation. The domain is different. The discipline is identical.
If you can manage a project, you can build an AI agent. But you have to do the architecture first. The planning phase isn't delay — it's the safety net that makes everything else possible.
Vibe coding without vibe architecting is dangerous. The architecture is what makes the coding safe.
───
What's next.
This is Episode 1 of an ongoing series. I'll be documenting Archie's development as he takes on more of Beyond Arc's operations — from autonomous trading on his own funded account, to content creation, to building revenue-generating products.
The next episode covers what happened when I went on holiday and left Archie running. Did the morning briefings keep coming? Did anything break? Did he do anything I didn't expect?
If you're considering building an AI agent for your business — or if you've been told you need to be a developer to do this — follow along. I'm proving that's not true, and I'm showing the working.
───
Adam Thorpe is the founder of Beyond Arc Ltd, operating across trading, technology, and consultancy. Archie is built on the OpenClaw framework using Anthropic's Claude as the primary LLM.
This series documents the real journey — wins, mistakes, and everything in between.
If this made you think differently about building AI agents, forward it to someone who needs to hear it.
Adam Thorpe
Founder, Beyond Arc