Reasoning, planning & the agent loop.

Week 3 was about one great prompt. This week the model gets to send many prompts — a loop where it sees a result, thinks, and chooses its own next step. The loop is the part of an agent that you own: you can read it, budget it, audit it, and stop it.

📌 This week's logistics: Group Assignment 1 is due — Tue Sept 15 (Wed section) / Wed Sept 16 (Thu section), 11:59 pm EST, on iCollege (enrollment & submission details). The final-project proposal is released on iCollege this week. And the Agent Radar presenter slot for next week is open — sign up.

The big question

Discussion questionIn Week 2 you watched Codex loop: observe → decide → act, over and over. Who decides that the loop should run one more time — and who decides that it should stop?

The model proposes ("I'm not done — the tests still fail"), but the harness disposes: the loop code you write decides whether to grant another iteration, checking the stop condition, the step cap, and the budget. An agent that "knows when to stop" is really a loop that checks. This week you build that loop yourself — first on paper, then in LangGraph.

Why one prompt is not enough

One shot vs the loop

One-shot prompt

Q: Which is taller — the tallest building in Atlanta, or twice the height of the Statue of Liberty? Model: (from memory, no checking) "Twice the Statue of Liberty is taller." ✗

No tools, no verification, no recovery. If the memory is wrong, the answer is wrong — and nobody notices.

The loop

Thought: I need both heights. Search first. Action: Search("tallest building Atlanta") Observation: Bank of America Plaza — 312 m Thought: Now the statue, then compare. Action: Search → 93 m · Calculator → 186 m Thought: 312 > 186. Done. Answer: the Atlanta building — by 126 m. ✓

Each observation is fed back before the next decision. Wrong turns become visible — and recoverable.

The trade: every extra loop step is another model call — more latency, more tokens, more money. This week is about buying reliability with steps, and knowing when you're overpaying.

This week's pages

Work through in order

01 The agent loopReactive vs goal-oriented · the Observe → Reason → Act cycle · step through a real ReAct episode · where the human gate goes. → 02 Reasoning patternsReAct · plan-then-execute · reflection · Tree of Thoughts — what each buys, what each costs, and how to choose (with the papers). → 03 LangGraph — the loop you ownState, nodes, edges · the canonical agent graph, clickable · build a minimal agent on the Gemini free tier · the max-iteration guard. → 04 Practice & caseAnnotate a real trace · draw the state graph · the itinerary case (greedy vs plan-and-check) · interview check · readings. →

Why managers care

The loop is the auditable part

Auditability

An explicit loop logs every thought, tool call, and state change. When compliance asks "why did the system do that?", you replay the trace.

Cost control

Steps are the unit of spend. A step cap and a token budget turn "AI costs" from a surprise into a line item.

Reliability

Feeding observations back is what lets an agent notice failure and recover — the difference between confident nonsense and a checked answer.