AI Agents: Conceptual Reflection

1. Choose one of the five topics below.
2. Write a focused essay of 2–3 pages (12pt, double spaced).
3. State your position or argument clearly in the first paragraph — do not just summarise, take a stance and defend it.
4. Ground your argument in concepts from class: PEAS, agent types, the ReAct loop, single vs multi-agent systems, tool use, or the agent loop.
5. You may cite external sources but it is not required — strong reasoning from first principles is equally valued.
6. Submit as a PDF.

Topics (pick one):

1 — PEAS in the Wild

1–2 pages

Find any real deployed AI agent — a customer service bot, a coding assistant, a recommendation system, anything you interact with in daily life. Map it rigorously onto the PEAS framework. The interesting part is not filling in the table — it is identifying what the designers chose to leave out of the performance measure and arguing whether that omission was intentional, negligent, or unavoidable.

2 — The Fifth Type

2–3 pages

Russell and Norvig describe five canonical agent types ending with the Learning Agent. Argue whether a modern LLM-based ReAct agent like the one we built fits cleanly into one of those five types, spans multiple types simultaneously, or represents something new that their taxonomy did not anticipate. Take a position and defend it with specific reference to the agent's components — do not just describe, argue.

3 — When Should the Agent Stop and Ask?

2–3 pages

Our agent runs autonomously until it produces a Final Answer. But there are situations where the right action is to pause and ask the user for clarification rather than guess. Propose a concrete decision rule — not vague guidelines but a precise condition — that determines when an agent should interrupt its loop and ask a follow-up question. Then identify two scenarios where your rule works well and one where it fails, and discuss what that failure reveals about the limits of rule-based autonomy.

4 — One Observation, Two Agents

2–3 pages

Imagine the same observation is returned to two different agent types — for example, a web search returns contradictory information from two sources. Trace in detail how a Goal-Based Agent and a Utility-Based Agent would each process and act on that observation differently. What does the difference reveal about what each architecture actually optimises for, and which one handles uncertainty more honestly?

5 — The Router Is the Agent

2–3 pages

In a multi-agent system, the router/orchestrator does not answer questions — it decides who should. Argue whether the router itself qualifies as an agent under the definitions we studied, or whether it is better described as something else — a dispatcher, a classifier, a meta-agent. Your answer matters because it determines whether the router should be held to the same design and safety standards as the agents it controls.