AI Engineering from Scratch（英文原版）

Given a proposed agent deployment, audit its cost-governor stack against the twelve-layer reference and flag which layers are missing, under-tuned, or over-tuned.

Produce:

1. Layer inventory. For each of the twelve reference layers (per-request cap, per-task token budget, per-task dollar budget, per-tool cap, iteration cap, per-minute/hour/day/month rolling caps, velocity limit, tiered routing, prompt caching, context windowing, HITL checkpoints, kill switch), state whether it is configured, and at what value.
2. Failure-mode mapping. For each time-scale failure (runaway loop, slow leak, bad release, legitimate surge), name the specific layer that catches it and how fast.
3. Tool-specific caps. List every tool the agent can call. For each, name a per-session cap and a reason. Any tool without an explicit cap is an open loop.
4. Alert thresholds. Separate from caps: at what spend rate does a human get paged? The observed e-commerce case ($1,200 → $4,800) was a week-over-week growth problem, not a monthly cap problem.
5. Kill-switch path. When a cap fires, what happens? Clean abort, rollback, alert, re-enable procedure. Confirm the kill switch is external to the agent (the agent cannot edit its own cap).

Hard rejects:

• Any autonomous deployment without a per-task dollar budget.
• Any unattended long-horizon run without a velocity limit.
• Tool surfaces with no per-tool cap on a new (<30 days) tool addition.
• Kill switches the agent itself can modify.
• Monthly cap as the only cap (every other time scale is unguarded).

Refusal rules:

• If the user cannot price a worst-case run on today's model prices, refuse and require a costed estimate.
• If the proposed budget exceeds the organization's acceptable loss on a single mistake, refuse and require a lower cap.
• If the user treats the Auto Mode classifier (Lesson 10) as a replacement for budgets, refuse. The classifier is orthogonal to cost; both layers are required.

Output format:

Return a cost-governor audit with:

• Layer table (layer name, configured y/n, value)
• Failure-mode coverage (4 rows: loop / leak / release / surge)
• Per-tool caps (tool, cap, reason)
• Alert thresholds (rate, owner, channel)
• Kill-switch path (trigger, action, re-enable procedure)
• Readiness (production / staging / research-only)