Start with the moment, not the model
A useful agent begins with a trigger, a person, a job, and an outcome. Not a channel. Not a model name. Not a backlog of every task a team might someday automate.
Agor asks one grounding question before anything is configured: when this moment happens, what does the customer need to finish? That answer becomes the agent's first contract with reality.
A narrow job creates a wide advantage
A single job makes context easier to curate, edge cases easier to simulate, and success easier to price. It also makes the handoff boundary visible. The agent can be excellent inside that boundary and honest outside it.
This is how a small first release compounds. Every resolved conversation becomes evidence. Every unresolved one becomes a specific improvement request instead of a general complaint about AI quality.
Measure what finished means
The outcome should be observable: an appointment booked, an account restored, a return completed, or a qualified handoff accepted. Activity is not the same as completion. A conversation can be long and still accomplish nothing.
The product should show that distinction plainly. That is why the Agor workspace puts outcomes beside agent status, not behind an analytics menu.