Subsumption Architecture is implemented by layers each of which implements a behavior and is comprised of one or more augmented finite state machine (AFSM). The bottom-most layer computes commands for the actuators based on immediate sensor readings. Higher layers look at the sensor readings and moderate the activity of lower ones by inhibiting or suppressing specific AFSMs in the lower levels.
Press here for a more detailed look at the architecture.
Subsumption Architectures may be considered modular because each AFSM performs an independent task. They are layered because that is the way the AFSMs are organized: different levels control different aspects of the architecture's behavior. Lastly, they are hierarchical, which is a stronger form of being layered. AFSMs of higher levels may moderate the behavior of ones below by inhibition or suppression. This gives each level its own "rank of control", creating a hierarchy.
Go to a discussion of this property for multiple architectures.