Month: August 2026

A fixed tool catalogue is a guess about which operations matter, made at design time by someone who couldn’t have known what the real problem would look like. The model can select from that guess. It can’t exceed it — and that ceiling, not the model, is what most agentic systems are actually hitting.

A declared project creates truths that are real, binding, and temporary. Store them as customer history and stale commitments contaminate later work. Promote them to firm doctrine and a one-site concession quietly becomes policy. An AI system can remember everything and still be operationally wrong.

Maximum reasoning effort doesn’t make an agent go further — it charges a premium on every step it was always going to take. Persistence is architectural. The dial is just a tax.

Aftermarket businesses think they sell availability. Most of them sell inventory plus an emergency reconstruction of the customer’s world, performed under pressure and billed to nobody. Preparedness is the part that could have been a product.

Switching observability off in development doesn’t save money — it moves the first honest look at your agent’s reasoning into production. With non-deterministic systems, the output tells you almost nothing about the path that produced it.

In bespoke consulting, the commitment gets priced at the moment both sides know least — and delivery pays to rediscover what the SOW pretended was settled. Using AI to write proposals faster doesn’t fix that. It just freezes weak assumptions with better formatting.

When capable people reopen ten systems every morning and call it process, they aren’t slow — they’re compensating for a missing layer. Most AI “productivity problems” are continuity-design failures wearing a costume.

The strongest evidence in an AI evaluation isn’t the answer — it’s what the system does when you remove the ground beneath it. Manufacture that evidence and you’ve committed the exact failure you were testing for.

An expectation written down beforehand can be wrong. That is its entire value. Everything else is retrospective narration — and if you didn’t state the expectation first, you don’t have results, you have a story.

“Are we confident this will work?” is a question nobody can be wrong about, before or after. Naming what breaks first — and in what order — puts executive judgment on the record, which is precisely why most rooms avoid it.