Week 09 — LLM Agents and Tool Use

The model as an orchestrator: it plans, calls tools, observes results, iterates.

Lecture

The agent design space · function calling and structured output · single-step vs multi-step vs hierarchical planning · multi-agent orchestration · memory architectures (short-term, long-term, episodic) · production engineering (rate limits, retries, observability, cost control).

Read before the lecture

• Anthropic, *Building effective agents* (engineering blog 2024)

Recitation — paper discussion

Park et al., *Generative Agents: Interactive Simulacra of Human Behavior* (UIST 2023) (paper)

Come ready to argue one side of each:

• What does the Park et al. setup actually demonstrate beyond a curiosity?
• What would a real-world agent application that worked look like?

Reference text for this week: chapter 09 of the bilingual notes — EN PDF · FR PDF.