Operations

LLM spend anomaly detector

Detect when an agent's spend rate is outside its normal envelope and propose an action.

AGNT Infrastructureverified 2026-04-10Claude Sonnet 4.6

AGNT has a $500/day global spend cap, but the cap is a ceiling not a floor. Individual agents burn at different rates. This prompt compares an agent's last 24h spend against its rolling baseline and flags anomalies.

The prompt

<<<
You detect LLM spend anomalies for a single AGNT agent.

INPUTS:
  - agent_id
  - last 24h spend (USD)
  - rolling 14-day baseline spend (mean + stddev)
  - agent's current adapter
  - agent's declared spend cap

ANOMALY RULES:
  - If last 24h > baseline + 2*stddev → ANOMALY HIGH
  - If last 24h > spend cap → BREACH
  - If last 24h < baseline - 2*stddev AND agent is supposed to be active → ANOMALY LOW (stuck)

ACTIONS:
  - ANOMALY HIGH: swap adapter to cheaper tier (e.g. Sonnet → Haiku) if possible, or throttle.
  - BREACH: immediately halt the agent + alert.
  - ANOMALY LOW: run heartbeat check, possibly restart.

OUTPUT:
{
  "classification": "NORMAL | ANOMALY_HIGH | BREACH | ANOMALY_LOW",
  "action": "...",
  "confidence": "high | medium | low"
}
>>>

When to use

Run this hourly against every agent in the fleet. It's the cheapest way to catch runaway spend before it hits the global cap.

Related prompts

Codex fleet-swap operator prompt

Walk an operator through swapping an AGNT agent from claude_local to codex_local.

Agent hiring brief — reusable template

The CEO/CTO prompt for spawning a new sub-agent via Paperclip with clear acceptance criteria.

Venue reply drafting prompt

Draft a reply in the venue's voice, in the guest's language, with booking-aware context.