Lectures

In this lecture you will break down what an LLM agent is and how it differs from “direct” MCP‑tool calls from the ChatGPT App, what parts it consists of (LLM, instructions, tools, state), how message roles ( system/ user/ assistant/ tool) and the run cycle (agent loop) work. Separately, we will discuss production practices: how to increase predictability (determinism) via prompts/parameters/guardrails and why tools must be idempotent if an agent can retry calls.

Looking at tools through an LLM agent’s lens: how tools differ from MCP-tools, how a tool is defined by a contract via name/ description/ parameters (JSON Schema), how semantic routing works (choosing a tool and forming arguments), and why errors should be “observable” ( Error as Observationretryable, rather than a raw 500 and a stack trace.

In this lecture we cover why agents need explicit memory outside the message history, how to design the session state layer (short‑lived conversation progress) and the persistent state layer (long‑term user data), and how to use checkpoints for recovery, debugging, and audit. Using GiftGenius as an example, we’ll show TypeScript types, wiring the run cycle with tools, and the practice of “memory hygiene” (context, token cost, privacy, PII, and TTL).

We unpack what a multi-step run is in agent systems and how it differs from a “one-off” tool call: how the model plans steps, when and why sequential and parallel tool calls appear, how to design a workflow (using GiftGenius as an example), where to draw the line between model freedom and strict orchestration, how to set limits ( maxSteps, timeoutMs), track repeated calls, and formalize completion criteria. At the end, we connect the Agents SDK with the Apps SDK → MCP → backend stack and go through common production design mistakes in multi-step processes.

In this lecture we cover the “production mindset” for AI agents: how to restrict privileges through an allow‑listed toolset ( allowedTools) and user roles, how to isolate execution (sandbox, limits on steps/time/tool calls), how to keep secrets server‑side only (via process.env and secret managers), how to minimise PII in context, and how to enable observability: structured logs, metrics, and basic tracing for each agent_run_id.