Agent architecture map
One use case, 3 ways it could come to life. Each is a complete architecture, not a sketch. They differ in where commit authority sits and how much a person is asked to carry, which is the choice that actually matters.
Assistive
Agents gather and recommend. Every settlement is paid by a licensed adjuster.
Buys: A single reversible failure mode and the shortest path to production.
Gives up: Throughput. The adjuster headcount is the ceiling on the whole process.
Supervised
Agents settle inside a numeric limit. Anything above it waits for an adjuster.
Buys: Most of the throughput, while keeping a person on every consequential claim.
Gives up: A real staffing commitment to watching the queue, which is routinely promised and not funded.
Bounded autonomy
Agents settle a narrow, well-tested band end to end. Everything outside it escalates.
Buys: Speed on the routine majority, with a human cost that scales with exceptions rather than volume.
Gives up: The band has to be defined precisely and evaluated continuously, which is where this variant usually dies.
Cheek LLC · Agent architecture map · v1