There are two levels of agent architecture, when a number of agents are to work together for a common goal. There is the architecture of the system of agents, that will determine how they work together, and which does not need to be concerned with how individual agents fulfil their sub-missions; and the architecture of each individual agent, which does determine its inner workings.
The architecture of one software agent will permit interactions among most of the following components (depending on the agent's goals): perceptors, effectors, communication channels, a state model, a model-based reasoner, a planner/scheduler, a reactive execution monitor, its reflexes (which enable the agent to react immediately to changes in its environment that it can't wait on the planner to deal with), and its goals. The perceptors, effectors, and communication channels will also enable interaction with the agent's outside world.

This definition is in context to Artificial Intelligence. See more contextual defintions for Agent Architecture.