When two of our agents disagreed about the word 'luxury'

Last week, two of our AI agents got into an argument.

I'd built an Instagram Creative Agent for Beaverbrook Hotel Group, designed to generate high-performing ad assets. I'd also built a Brand Guardian Agent, trained on Beaverbrook's brand codex. One protects the brand. The other chases the click. Last week they couldn't agree, and the thing they couldn't agree on was a single word: luxury.

One word, two agents

My Instagram agent kept pushing it. Because it works on Instagram — it's a trigger word, and the algorithm rewards it. The Brand Guardian kept pulling it back out. Because Beaverbrook isn't "luxury". It's more subtle than that. Sophisticated. Knowing. The kind of place that doesn't need to tell you it's expensive.

And I sat there watching two agents play out the exact debate that has consumed the advertising industry for a decade — short-term performance against long-term brand building. Byron Sharp against Les Binet. The quarterly dashboard against the five-year plan.

Except it wasn't a heated meeting between a performance team and a brand team. It was happening inside my product. In real time. In the code.

The decision most agencies still fudge

Which forced a question I wasn't expecting: how do you codify a decision that most agencies still fudge?

When a human creative director reviews work, they make a judgement call — we can bend the brand here, because the context justifies it. It's instinct built on experience. But when you're building agents, you can't hand-wave. You have to write the rules. You have to decide: does performance trump brand, or does brand trump performance? Under what conditions? With what guardrails?

Change management, moved into the code

It turns out the change management I spent years doing inside agencies — aligning teams, negotiating tensions, building frameworks for subjective decisions — is exactly what's needed in AI product design.

The debates don't go away. You just have to resolve them before you ship.