AI agents
1. Definition
An AI agent is an autonomous entity that perceives its environment through sensors, decides what to do using its own reasoning (a policy or model) and acts on that environment through actuators in order to achieve its goals (the classic definition from Russell and Norvig). An agent is not just a program that answers a question — it observes, decides and acts, repeatedly, in a loop, and its behaviour improves with feedback. Formally: Agent = Architecture + Program.
2. The Agent–Environment Loop
┌──────────────────────── ENVIRONMENT ────────────────────────┐
│ (world state: rooms, roads, market, web pages, game board) │
└───────▲──────────────────────────────────────┬─────────────┘
│ actions │ percepts
┌───────┴──────────────────────────────────────▼─────────────┐
│ AI AGENT │
│ SENSORS → PERCEPT → [ internal state / model ] │
│ ↓ │
│ REASONING / POLICY │
│ (rules, search, ML, LLM) │
│ ↓ │
│ ACTUATORS ← ACT ← [ goal / utility / plan ] │
└────────────────────────────────────────────────────────────┘
goals, prior knowledge, performance measure
3. Types of Agents
| Type | How it decides | Example |
|---|---|---|
| Simple reflex | Condition–action rules on the current percept only | Thermostat: if temp > 28 → switch AC on |
| Model-based reflex | Keeps an internal state/model of the world it cannot fully see | Robot vacuum remembering already-cleaned rooms |
| Goal-based | Chooses actions that move it towards a goal, using search/planning | GPS route planner, chess engine |
| Utility-based | Maximises a utility (happiness/score) when many goals conflict | Ride-hailing dispatch balancing wait time and cost |
| Learning agent | Has a learning element and a critic that improve the policy with experience | Recommendation engine, AlphaZero |
| Multi-agent system | Many agents coordinate, cooperate or compete | Traffic simulation, market bots, drone swarms |
Modern LLM-based agents add to the above: a language model as the reasoning core, tools/function calling (search, code, database, APIs), memory (vector store/RAG, conversation state), planning (task decomposition, ReAct loops) and reflection on results.
4. Properties of a Good Agent
Autonomy — acts without step-by-step human instruction · Reactivity — responds to changes in the environment · Pro-activeness — takes initiative towards goals rather than only reacting · Social ability — communicates and negotiates with other agents or people · Rationality — chooses the action that maximises expected performance · Adaptability/learning — improves from experience · Mobility (in some taxonomies) — can move across networks/hosts.
5. PEAS — Describing an Agent (exam favourite)
| Agent | Performance measure | Environment | Actuators | Sensors |
|---|---|---|---|---|
| Self-driving taxi | Safety, time, fuel, comfort, legality | Roads, traffic, pedestrians, weather | Steering, accelerator, brake, horn, display | Cameras, LIDAR, radar, GPS, speedometer |
| Medical diagnosis agent | Correct diagnosis, patient safety, cost | Hospital, patient, lab | Screen recommendations, order tests, alerts | Patient history, symptoms, test reports |
| Smart-home thermostat | Comfort, energy saving | Rooms, weather, occupancy | AC/geyser relay, fan speed, alert | Temperature, humidity, motion, door sensors |
| Chess-playing agent | Win the game, fast move | 64-square board, opponent | Move a piece (in a physical board: arm) | Board position, clock |
| Study/notes assistant (e.g. this app's AI) | Helpfulness, accuracy, syllabus coverage, safety | Question bank, syllabus, student profile | Reply text, links, generated notes/quizzes | User question, subject/topic context, past interactions |
6. How Modern AI Agents Are Built (2023 onwards)
- Reasoning core — a large language model (or a smaller task-specific model).
- Tools — web search, calculators, code execution, database/API access, file operations.
- Memory — vector databases (retrieval-augmented generation) for long-term knowledge, plus short-term conversation state.
- Planner — decompose a goal into steps: ReAct (reason + act), plan-and-execute, or tree search over plans.
- Reflection and self-check — the agent reviews its own output and retries on failure.
- Guardrails — permission scopes, allow-listed tools, rate limits, sandboxing, human approval for risky actions (payments, deletions), logging and auditability.
Examples in the wild: coding agents (write, run and fix code), customer-support agents, browser/computer-use agents, research assistants, RPA-style business process agents, and multi-agent frameworks (AutoGen, CrewAI, LangGraph) in which specialised agents delegate to one another. MCP (Model Context Protocol) is an emerging standard for exposing tools and data to agents in a uniform way.
7. Advantages and Limitations
| Advantages | Limitations |
|---|---|
| Automates multi-step work, not just single answers | Errors compound over a long chain of steps |
| Works 24×7, consistently, at low marginal cost | Can hallucinate or misuse a tool confidently |
| Can use many tools and data sources at once | Cost and latency grow with reasoning depth |
| Improves with feedback and data | Needs oversight, permissions and guardrails; risky actions must be gated |
| Scales expertise to many users | Prompt injection, data leakage, opacity of decisions |
| Frees humans for judgement and creativity | Accountability: who is responsible when an agent errs? |
8. Exam Points
- Define an AI agent and draw the agent–environment loop — 5 marks, almost certain.
- Explain the types of agents (simple reflex, model-based, goal-based, utility-based, learning) with examples — 5-7 marks, very frequent.
- Write the PEAS description for a given agent (taxi/thermostat/chess) — 5 marks, frequent.
- List the properties of an intelligent agent (autonomy, reactivity, pro-activeness, social ability) — 5 marks.
- What is a multi-agent system? — 3 marks.
- Explain how a modern LLM-based agent uses tools, memory and planning — 5 marks.
- What are the risks/limitations of autonomous agents? — 3-5 marks.