Part 9 of 11 · 1 min read
Quick notes
- Support agents are the most common production agent, handling routine requests round the clock and escalating the rest. Scope narrowly: a few intents with clear rules.
- The model reads and drafts; code decides. Policy (thresholds, refund limits, legal language to a human) lives in code, and tools enforce their own preconditions and take idempotency keys.
- Autonomy is a dial, not a switch. In the simulation, unguarded autonomy automated the most and refunded the most money wrongly; a confidence threshold and a refund cap cut wrong actions to almost nothing.
- The right autonomy depends on accuracy and on the cost of a mistake, so measure a real model's calibration and tune thresholds on real traces. Safety that lives in the tool survives a confused model.
- Escalation is a feature: a handoff packet means nobody re-asks the customer. Operations agents need risk tiers (read-only, mutating with approval, forbidden) and a capped write rate.