2. Core Mental Model

The central idea of the course is:

An LLM is a stateless reasoning engine. An agent is a stateful system built around that engine.

The LLM does not remember previous calls on its own. It does not run tools, enforce limits, or guarantee that it stops. The software system around it, the agent harness, supplies all of that.

Definition. The agent harness is the software runtime around the LLM engine that turns it into an agent. It keeps the state the model does not have. It builds the context for each call and keeps it within limits. It runs tool actions under explicit permissions. It drives the loop and decides when to stop. It records what happened so you can debug and evaluate. The model is a part of the harness, not the other way around.

A useful operating-system analogy:

Component	Computer Analogy	Responsibility
LLM engine	CPU	Runs computation on the data it is given
Context window	CPU registers	The only state the model can see during one call
Agent harness	Operating system	Manages state, I/O, permissions, scheduling, recovery, and observability
Tools	Devices / system calls	Capabilities that cause real side effects
Guardrails	MMU - virtual memory (protection between processes)	Deterministic checks that limit what may happen
Traces	System telemetry	Records what happened for debugging and evaluation

This slogan states the design principle:

Deterministic harnesses for non-deterministic brains.

The model's output is probabilistic, so the harness must compensate with deterministic control: it manages state explicitly, keeps the loop bounded, validates output, and handles errors in a predictable way. That is what "deterministic harnesses for non-deterministic brains" means in practice.