Kurnell plans products alongside AI agents. We think that sentence should be a far more concrete commitment than "we use AI," so this first post sets out exactly where we drew the line between people and agents.
What people hold
Direction and standards. Which problem to solve, why now, and when it is good enough. We do not delegate those three. Most of the context behind such judgements lives outside any document, and people are the ones who pay when the call is wrong.
What agents take on
The repetitive, labour-heavy stretches: research, requirements, prototypes, and checking results against the rules. These converge on a verifiable answer, which makes the cycle far shorter than doing it by hand.
What the boundary costs
The failure we meet most often is not an agent producing a wrong answer — it is an agent producing a plausible answer with no evidence behind it. So we look at the evidence before the output. If a judgement cannot be traced to a specific file and line, it does not pass. That single rule trims a little speed and saves a great deal of rework.
What comes next here
Notes on technology and marketing, gathered while planning and verifying products. We intend to write up the hypotheses we abandoned and why, not only the decisions that worked. A record of decisions becoming the starting point for the next plan is what we mean by compounding knowledge.