The Orchestrator

When goals are foggy, AI does not wait for clarity. It fills the fog with features.

I named this pattern the Scope Creep Kraken in The AI Developer’s Field Guide because teams needed a word they could say out loud in a standup without writing a white paper.⁹ Strip away the sea monster and the failure is entirely ordinary: AI collapses the cost of proposing and prototyping, so “we can generate this quickly” quietly replaces “we decided this belongs,” and scope grows tentacle by tentacle with each addition locally reasonable. The two-person tool becomes a platform with seven tracks and no additional people, sprint metrics look incredible, and validation against the original goal disappears somewhere around week three.

The first tentacle is usually sincere. Why not also render it on mobile — ten minutes? Then Kafka validation. Then Terraform. Then a web UI. Then a roadmap that needs a program manager the company never hired. Nobody felt reckless at any point, which is the whole point: the Kraken does not require incompetence, only cheap generation and unclear exclusions.

Which is where incentives decide the profession. If the incentive is velocity of shipping features, the system optimizes for commits, pull requests, options, and demos; if the incentive is a measurable business KPI — activation, refund completion time, claim accuracy, onboarding success — then the system finally has something it can refuse. An orchestrator exceeds a developer here not by typing less but by building logic that focuses on outcomes rather than code. The developer’s loop asks whether the change compiles and passes tests. The orchestrator’s loop asks whether the change moves a named outcome without crossing a named boundary, and whether the next feature is an answer or a tentacle.

You can feel that difference in a planning meeting. A feature-velocity room leaves with fourteen polished options and a warm sense of productivity, while an outcome room leaves with three options, two exclusions, and one metric that would make the work a failure even if the code is beautiful. The second meeting is less exciting, and it is the one that keeps people from drowning. Goal clarity is not a kickoff slide but an operating control: write the exclusions, write the KPI, write what done means when generation is cheap.