Nanyuki is not Silicon Valley. It's a town in central Kenya where the Wi-Fi is intermittent, the power sometimes cuts without warning, and the nearest tech conference is a five-hour drive away. From a Mac in that town, we've been running 12 AI agents that power real client work for an accounting firm and a gym. We didn't plan this. We tinkered our way here. And somewhere between the third agent and the twelfth, I stopped thinking of this as an experiment and started thinking of it as infrastructure.
This is what actually happens when you try to build at scale without scale. Not theory. Not a playbook copied from San Francisco. Just what broke, what we fixed, and what we learned about systems that run without you.
The Tinkerer's Path to Complexity
Isaac doesn't do five-year plans. He builds things. He writes code, deploys it, watches it fail, rewrites it. This is how we ended up with 12 agents instead of the one or two you'd expect.
It started with Jack, the orchestration agent. Then Kimani came in for engineering decisions. Wanjiru handles backend, Baraka owns frontend, Amara runs DevOps on GCP. Taini does QA. Mwangi manages finance. Helen handles communications. Eric runs gym operations. Mandi does research and business development. Safi owns security.
Each agent has a name, a role, and a domain where they own decisions. Not ownership in the corporate email-chains sense. Real ownership. When something breaks in their domain, they fix it without asking for approval. They know the context. They know why decisions were made.
Ownership Beats Simplicity
The mistake everyone makes is thinking complex systems fail because they're too complex. Wrong. They fail because nobody knows who's responsible when things go sideways.
We have 12 agents. That's complex on paper. But each agent knows exactly what they own. Amara knows the infrastructure. If a deployment fails, she knows it's her domain. She doesn't need a meeting. She doesn't need to wait for someone else to decide. She moves.
This is radical compared to how most companies work. Most places add layers of coordination because they're terrified of chaos. More meetings, more Slack channels, more approval gates. The result is that nothing moves fast and nobody's sure who failed.
We chose the opposite. Clear domain boundaries and fast ownership beats trying to eliminate chaos. You can't eliminate it anyway. You can only design for it.
The Agent Command Centre Keeps Everyone Honest
We built a live dashboard. It shows every agent currently active, what they're working on, activity logs, team chat, finance flows, security posture. This one tool changed how we think about running 12 agents.
It's not a status report system. It's a transparency system. Everyone can see everyone else. When Helen in comms sees Wanjiru's backend agent is working on something, she can understand the impact on client communication. When Amara's DevOps work is running, Taini in QA sees it and can think about what needs testing.
No meetings to sync this information. No email threads. Just visibility. The other thing it does is make failure visible immediately. When something breaks, it doesn't stay broken in some hidden subsystem. It's right there on the dashboard.
What This Actually Powers
None of this matters if it's just theory. We run this system on real work. Petrus CRM for an accounting firm is a KES 1.3 million contract. That's not toy money. It needs to work. When Petrus handles financial records for a professional firm, there's no room for iterating in production.
We also power Nanyuki Active, a gym in Nanyuki. Every member managed, every billing cycle, every class booking depends on these agents working together. When the system fails, people can't access the gym they pay for.
This is where the theoretical breaks down and the practical takes over. We're not writing a proof of concept. We're running infrastructure for actual businesses in Kenya. The fact that it works, and keeps working, tells us something important. You don't need to eliminate complexity to ship real systems. You need to own it.
Agentic Workflows Are About Designing Systems That Run Without You
The core insight is this: agentic workflows aren't about replacing people. They're about designing systems that don't require constant human attention to move forward.
When this started, every decision ran through Isaac. Then it didn't. The agents started handling things. Not perfectly, not without bugs, but they handled things. The system runs and Isaac isn't a bottleneck anymore.
That's the point. Your system shouldn't require you to be there. The moment you become critical infrastructure, you've failed at building systems.
We designed clear interfaces between agents. Jack orchestrates but doesn't dictate. Kimani makes engineering calls without waiting for approval. Safi owns security and moves when something looks off. Each agent operates independently in their domain while contributing to the whole. This is what scales. Not hiring more people. Building systems that move without you.
Messiness Is the Feature, Not the Bug
We could pretend we built a perfect system. We didn't. There's messiness everywhere. Agents sometimes retry things they shouldn't. Context gets confused. Decisions conflict and we patch them after the fact.
We've made peace with that. Most companies respond to mess by adding more process, more gates, more review cycles. That slows everything to a crawl.
We responded by designing clear interfaces. If the interface is clear, the system can be messy behind it. Baraka's frontend doesn't care about Wanjiru's backend chaos as long as the API contract holds. Mandi can run research loops without worrying about how Safi's security checks work internally. The interfaces are the boundaries.
That's the actual skill. Not building perfect systems. Building systems where clear boundaries let you tolerate a lot of internal mess.
This is what AI infrastructure actually looks like for a small firm in Kenya. Not polished. Not planned three years in advance. Built, broken, fixed, shipped. Running on open source tools and Claude. From a town where the Wi-Fi cuts out sometimes and people still get their work done.
Want to discuss AI for your business?
Let's talk about how custom software can transform your operations.