Agent harness
The runtime scaffolding that provides tools, permissions, context, events, limits, and recovery around an AI agent.
Agent loop
A repeated cycle in which an agent observes state, chooses an action, executes it, evaluates the result, and decides whether to continue.
Checkpoint
A durable snapshot of state and artifacts that lets a loop recover without repeating all prior work.
Context engineering
The discipline of selecting, structuring, refreshing, and isolating the information available to a model at a given step.
Evaluator
A deterministic check, model-based judge, human reviewer, or combined system that assesses an iteration against explicit criteria.
Guardrail
A technical or procedural boundary that prevents, validates, or escalates risky agent behavior.
Human-in-the-loop
A design in which a person supplies judgment, approval, correction, or escalation at defined points in an automated workflow.
Loop budget
The allowed spend of iterations, tokens, tool calls, time, money, or risk before a loop must stop or escalate.
Loop Engineering
The discipline of designing the system that repeatedly guides, observes, evaluates, and constrains an AI agent.
Loop Engineer
An emerging role focused on the objectives, context, harness, feedback, stop conditions, observability, and governance around agent loops.
Stop condition
A rule that ends or escalates a loop after success, failure, budget exhaustion, timeout, unsafe state, or insufficient progress.
Tool boundary
The permissions and validation contract that determines which external actions an agent may perform and under what conditions.