Notation & glossary
The series is deliberate about symbols because later essays inherit them. Defined once in The Harness Is the Product, reused throughout.
Symbols
MThe model — the weights θ. One block inside a larger system.
HThe harness — the six components that wrap the model into an agent.
EThe environment / tool-world the agent acts in.
τA trajectory — the multi-step episode the agent runs.
s, a, o, rState, action, observation, reward at each step of the loop.
ΣThe skill / rule store — memory that outlives a single episode.
C_fbEffective feedback compute — the verified, non-redundant, retained feedback the harness extracts per task. The harness scaling coordinate.
The six harness components
The harness H is the closed loop that turns a stateless next-token predictor into an agent that acts over many steps. Strip it to its irreducible parts and six show up every time — the decomposition the series reuses.
Context assemblyWhat enters the prompt each turn: instructions, retrieved facts, prior observations, trimmed to fit.
Tool routingChoosing which tool or model call to make, and how the result is delivered back.
State & memory (Σ)What is retained across steps and episodes: verified facts, failed attempts, constraints.
VerificationChecking whether an action's result is correct or sufficient before the agent builds on it.
RevisionWhat the agent does when verification fails: retry, repair, replan.
OrchestrationThe controller that sequences all of the above and decides when to stop. Owns C_fb.
The loop
A trajectory τ is a sequence of steps: from state s the model emits an action a (a tool call or a token), the environment E returns an observation o and a reward r, and the harness updates its state and memory Σ before the next step. The loop — not the single forward pass — is the unit of analysis, and C_fb measures how much usable feedback it extracts per task.