There is a temptation in agent design to give a system as much capability as the medium allows. The model can browse, retrieve, plan, write, ship. So the agent does all five. The result is a system that demonstrates everything and is reliable about none of it. It cannot be evaluated. It cannot be observed. It cannot compose with anything else.
The studio's working position is that an agent is not a service. An agent is a contract. The contract has three terms — what the agent accepts as input, what it refuses, and what it returns as output. Those three lines define the shape. Everything else — model choice, prompt design, tool surface — is implementation under the contract.
A contract that says 'this agent reviews vendor invoices, refuses anything that is not a vendor invoice, returns a structured approval or a structured rejection with reason' is a system. A contract that says 'this agent handles vendor workflows' is a vibe. The first composes with five other contracts into an accounts-payable architecture. The second composes with nothing.
A well-designed agent refuses more work than it accepts.
The hardest part of agent design is not capability — it is refusal. A well-designed agent refuses more work than it accepts. It refuses input shapes it cannot reason about. It refuses ambiguity it cannot reduce. It refuses to improvise at the edge of its competence. Every refusal is the contract holding under pressure. Without the refusals, the contract is not real.
The studio's evidence from the field: narrow agents compose into large systems with shape. Broad agents collapse into single-system surfaces with no shape. The first compounds. The second drifts. This is true regardless of the model under the hood. The contract is the architecture; the model is the implementation.
An agent is a contract. The contract is the architecture. The architecture is what compounds. Most production AI systems being shipped right now are violations of this — they have capability without contract, and they cannot be made to hold. The studio's bet is that the next decade of intelligent systems will be defined by how well the contracts hold under pressure they were not originally written for.