Choose the smallest useful system.
More agents, a new dashboard or a larger model are implementation choices. The useful question is which design completes the job well enough to operate.
Architecture follows the task
Deploy Co.’s architecture heuristics favor explicit software for decisions that can be expressed clearly, with flexible model reasoning used where the task needs it. Anthropic’s guidance on effective agents likewise recommends starting with a simple design and adding complexity when it improves the outcome.
This is a decision to test. A workflow with stable rules may need routing and a few model calls. An open-ended investigation may need planning and iteration. Both need a clear finish and a way to inspect the result.
An existing work surface can be enough
OpenAI’s Plex Coffee story starts with company knowledge in Notion and an onboarding handbook, made easier for staff to use. Its experiments with supplier ordering are a separate, developing use case.
That distinction matters in discovery. First assess what the business already owns, which connections are available and where people need the result. Build a new interface when it solves a real operating problem.
Leave room to replace the design
A model, integration or standard product may make custom work unnecessary later. Keep source context, tests and operating assumptions clear enough to evaluate that change.
Our aim is a useful capability the team can maintain. Reusable components should make the next improvement easier while preserving the business’s own rules and context.
The sources describe the publishers’ own customers and systems. The delivery recommendations are our interpretation; they are not customer results or endorsements of this studio.