Key Components of an AI Agent
Every AI agent, regardless of complexity, is built from four fundamental components.
1. Sensors (Perception)
How the agent observes its environment.
- Physical sensors: Cameras, microphones, LIDAR, temperature sensors
- Digital sensors: API inputs, text input, database queries, web scraping
- Perception pipeline: Raw data → processed information the agent can reason about
2. Actuators (Actions)
How the agent affects its environment.
- Physical actuators: Motors, speakers, displays, robotic arms
- Digital actuators: Sending messages, API calls, writing files, executing code
- The set of all possible actions is called the action space
3. Environment
The world in which the agent operates. Environments are classified by:
| Property | Options | Example |
| Observability | Fully / Partially observable | Chess (full) vs. Poker (partial) |
| Determinism | Deterministic / Stochastic | Calculator (det.) vs. Stock market (stoch.) |
| Dynamics | Static / Dynamic | Crossword (static) vs. Traffic (dynamic) |
| Agents | Single / Multi-agent | Solitaire (single) vs. Auction (multi) |
| Continuity | Discrete / Continuous | Board game (discrete) vs. Driving (continuous) |
4. Performance Measure
How we evaluate whether the agent is doing a good job.
- Must be defined externally (not by the agent itself)
- Should capture what we actually want, not just a proxy
- Examples: accuracy, response time, user satisfaction, profit
The PEAS Framework
A useful framework for describing any agent:
| Component | Question |
| Performance | How do we measure success? |
| Environment | Where does the agent operate? |
| Actuators | What actions can it take? |
| Sensors | What can it perceive? |
Example — Self-driving car:
- P: Safety, arrival time, comfort, fuel efficiency
- E: Roads, traffic, pedestrians, weather
- A: Steering, acceleration, braking, signaling
- S: Cameras, LIDAR, GPS, speedometer
Key Takeaways
- All agents share the same four components: sensors, actuators, environment, performance measure
- The PEAS framework helps systematically design agents
- Environment properties determine which agent architecture is appropriate