An agent encapsulates a Sequencer, Driver and Monitor into a single entity by instantiating and connecting the components together via TLM interfaces. Since UVM is all about configurability, an agent can also have configuration options like the type of agent (active/passive), knobs to turn on features such as functional coverage, and other similar parameters.
Usually, it makes sense to create an agent that provides protocol specific tasks to generate transactions, check the results and perform coverage. For example, an agent can be created for the WishBone protocol whose sequencer will generate data items which can be sent to the driver. The driver then converts the data item class object into actual pin level signals and drive them to the DUT. The monitor may passively collect the outputs from the DUT, convert them back into another data item class object and distribute it among all the components in the testbench waiting for the item.