The Orchestrator

A lot of people focus on increased productivity, and very few stop to ask whether what is being achieved was worth achieving.

Yes, a team can create a new product that does X in a week of agentic magic. Is it the right thing to build? Is it worth ten thousand dollars in token costs, or a hundred thousand once marketing, support, and longer contexts join the party? AI spend feels casual right up until adoption, context growth, and concurrency turn it into a structural commitment, and organizations track story points and tokens while skipping purpose entirely: whether the work deserved to exist, what simpler tool it displaced, and whether anyone forecast the cost of success.

I have watched rooms celebrate a prototype that cost less than a dinner and then, three months later, treat a five-figure monthly bill as an entitlement because everyone depends on it now. That dependency was never designed; it accumulated. Nobody evil did this — a sequence of reasonable people said yes to useful-looking things — and that is how cost singularities actually form, not as a villain’s plot but as a failure to keep purpose attached to spend while the spend still looked optional.

Klarna’s correction had a cost story inside it — the chief executive’s own words were that cost had been “a too predominant evaluation factor” — and quality was the bill that arrived later. The language shifts from “is this worth doing?” to “how do we protect what we’ve built?”, and the person who keeps asking about cost starts to sound difficult. That social punishment is part of the antipattern, because scrutiny is not disloyalty.

There is a second cost trap that arrives dressed as virtue: endless optimization of unit costs after the big knobs have already been turned. The meeting lasts an hour, the follow-up lasts two weeks, the savings are a few hundred dollars a month, and the opportunity cost — the product decision nobody made because the team was counting beads — never appears on any dashboard. Orchestrators need both instincts at once: refuse spend without purpose, and refuse thrift that displaces judgment.

So the job includes unit economics that can survive a skeptical adult. Not only how many tokens per task, but what outcome this spend is for and what would make us stop. A system that can generate endlessly without a purpose check will generate endlessly, which is why purpose functions as a permission envelope for ambition.