Human in The Loop Gates

The industry has settled the debate: practitioners want calibrated autonomy — agents that act freely inside boundaries and stop at the right moments — not full automation and not a human babysitting every step. But "add a human in the loop" is not a design; it is a wish. This skill turns it into engineering: a tiered gate system that classifies every agent action by reversibility and blast radius, routes each tier to the right approval mechanism (auto-approve, async review, blocking sign-off), handles interrupt/resume cleanly, and produces the audit trail compliance actually accepts.

To design and specify the human-oversight layer for an agent system: which actions need approval and which do not (with written rationale), how approvals are requested and recorded, what happens when the human is unavailable (escalation, timeouts, safe defaults), how a human interrupts and resumes a run, and what the audit trail contains. The output is a gate specification the team can implement, plus the policy document that justifies every line of it.

• When launching any agent that takes real-world actions (sends email, moves money, changes production, contacts customers).

• When an agent incident has happened ("it sent the wrong email to 400 people") and you need gates that would have prevented the class, not just the instance.

• When compliance, legal, or a customer asks "who approved that?" — the skill designs the audit trail that answers it.

• When tuning an existing agent that pages humans too often (alert fatigue) or too rarely (silent failures).

• For purely read-only / analytical agents with no side effects — gates on reads are theater; spend the effort on output review instead.

• When the "human" is fictional (no one will actually review) — a gate nobody watches is worse than no gate, because it creates false assurance. The skill will say so.

• For split-second trading or real-time control systems — human gates add latency that breaks the loop; those need pre-authorized policies, not approvals.

• As a substitute for making the agent itself more reliable — gates catch failures; they do not fix an agent that fails 30% of the time. Fix the agent too.