A guardrail is a deterministic check outside the model that constrains what a run may do: an allowlist, a spend ceiling, a confirmation gate, an output filter. Its value comes precisely from not being a model, because it cannot be argued with.

Guardrail

Insert a human before irreversible or expensive actions, not after every step. Approval fatigue is real: an agent that asks twenty times per run gets rubber-stamped, which is worse than an agent that asks once at the point that matters.

Where to Put the Human

A topic of the Building AI Agents course. Planned for 1.5 hours at Medium difficulty.

University of Michigan - Ann Arbor

A category of the Building AI Agents course, covering 4 topics.

Planning, Orchestration and Evaluation

A topic of the Building AI Agents course. Planned for 2.5 hours at Hard difficulty.

Planning Strategies: ReAct and Plan-then-Execute

A topic of the Building AI Agents course. Planned for 2 hours at Hard difficulty.

Learn Before

Related

Learn After