Managing AI Agents
What building an 11-skill AI team taught me.
The Evolution. From Chatbot to AI Team
This website is managed by a team of 11 AI agents. Not one prompt. A structured team with defined roles, workflows, and quality checks. Here’s how it started.
In 2023, I built Chat GPSteve, an AI chatbot trained on my career history. It was an experiment in whether AI could represent me.
Two years later, I asked a harder question: could AI actually work with me? Not a single tool doing everything, but a structured team, specialists with defined roles, a clear workflow, and quality checks at every step.
Chat GPSteve
One chatbot. Representation.
The AI Team
11 specialists. Collaboration.
A System, Not a Tool
Using Claude’s Cowork platform, I built a system of 11 specialised AI skills that manage a website and content. Each skill has a single job, a defined scope, and knows how to hand off to the next.
There’s a content reviewer, a 358-line guide, A GEO optimiser that audits each page for AI search visibility. An SEO checker handles the traditional plumbing: title tags, heading hierarchy, and internal links.
A site planner acts as the project manager. Tracking which pages have been reviewed and recommending what to audit next. A reporter consolidates every finding into a single prioritised action plan. And a session manager ensures nothing gets lost between working sessions.
It isn’t one prompt trying to do everything. It’s an AI team with roles, a workflow, guardrails and feedback loops built in by human design.
Running Reviews Like Experiments
The order isn’t random, it’s by design. Content quality has to come first because everything else builds on it. AI visibility comes second because it depends on solid content underneath. Traditional SEO comes last, so it’s working with the final copy, not a draft that’s about to change.
It’s the same principle I learned running experimentation programmes: sequence creates compounding quality. You don’t optimise for conversions before you’ve understood what users are actually trying to do.
After all three have run, the reporter pulls everything together into one prioritised action plan, scored by impact and effort, so I know exactly what to fix first.
Quality of the underlying writing — clarity, accuracy, voice, value to the reader.
Audits each page for visibility in AI search — structure, entities, claims, sources.
Title tags, heading hierarchy, internal links — the traditional plumbing, last.
Consolidates every finding into one action plan, scored by impact and effort.
358 Lines of Guardrails
The most important part of this system isn’t the AI. It’s the constraints you put around it.
I wrote a 358-line guide that defines exactly how the system should work. Every skill references this core guide. The AI doesn’t guess what it should do. It has a specification.
At the centre of the system is human review. Nothing ships without a human reviewing it. The AI proposes. I decide.
What Managing AI Agents Actually Taught Me
Four observations from running the system for long enough to trust it — and to see where its sharp edges are.
Specialisation beats generalism.
One focused AI skill with a clear brief consistently outperforms a single prompt trying to do everything.
Sequence is a design decision.
The order skills run in isn’t random, it’s by design. Getting it wrong means that later skills are working with unstable inputs.
Memory changes everything.
An AI that remembers what happened last session is fundamentally different from one that starts fresh every time.
The human role evolves, it doesn’t disappear.
I spend less time writing first drafts and more time reviewing, directing, and refining. That’s not a reduction in human involvement, it’s an elevation of it.
This Is What I Help Organisations Build
The website AI team is a working prototype of what I help organisations design at a larger scale.
The principles are the same: define clear roles, build in quality frameworks, get the sequence right, maintain context between sessions, and keep a human in the centre making the decisions that matter.
I spent over a decade learning these principles through experimentation. Now I apply them to help organisations work with AI in a way that’s structured, measurable, and centred on the business.