Glossary

Terms used across the handbook. Each term is a unique anchor you can link to.

Loop engineering

Designing the outer loop an agent runs in — prompt, context, harness, and the loop itself — not just a single prompt. See what is loop engineering.

/goal

A primitive that runs depth-first toward a single verifiable condition, then stops, with a separate evaluator checking the condition each pass.

/loop

A primitive that runs on an interval with no built-in end — a cron in your terminal. Safe only when governed (stop condition + cost cap).

Loop contract

The six-field template (Goal / Context / Constraints / Done-when / Evidence / If-blocked) that specifies a governed loop. See the loop contract.

Maker-checker

Separating the agent that produces work from an independent evaluator that accepts or rejects it on executable evidence (writer ≠ checker).

Done-when

The single, verifiable stop condition. A separate checker — not the model's confidence — decides whether it is met.

Durability ladder

The progression from a one-shot prompt to a fully unattended automation, ordered by how little babysitting each needs. See /goal and /loop basics.

Context rot

The degradation of an agent's reliability as a single transcript grows. Loops avoid it by starting each pass from fresh context plus a small memory file.

LOOP.md

A small, durable memory spine file a loop writes/reads across passes so a fresh-context pass knows what earlier passes already did.

Circuit-breaker

A governance control that halts a loop after N consecutive no-progress or failing passes, so a stuck loop stops instead of burning budget.

Cost cap

A hard $/run ceiling that stops a loop regardless of progress — the difference between "run until done" and "run until bankrupt."

Back to the guide index.