Module II — Harness Anatomy

Status: outline. Lecture body not authored.

Full dissection of an agent harness:

Observation
Context
Control Loop
Tools
Memory
State
Verification
Execution

Study: context manager; tool registry; execution loop; state machine; memory; sandbox; retries; termination; checkpoints; evaluator.

Boundary Primitives: Context, Memory, State & Evidence

The harness must separate subsystems that are commonly conflated:

Context: the working material assembled and bounded for the current inference window.
Memory: retained information, experiences, or knowledge carried across inferences or sessions.
State: the deterministic, operational condition and execution phase of the system at a specific point in time.
Evidence: verifiable artifacts, execution logs, diffs, or test observations that substantiate a claim or decision.

Core Architectural Question:
Why are Memory and State separate subsystems?

*(The rigorous answer is developed in Module III and Module X: Memory informs; State governs.)*

Build a Minimal Agent Harness.