The GUARDIAN Framework: Production AI Agent Monitoring, Debugging, and Cost Control

*The operations manual for teams running LLM agents in production. Not "build your first agent." This is what you need after deploy day, when the agent is live, the bill is climbing, and someone is paging you about a hallucinated refund.*

There is no shortage of content on building agents. There is almost nothing on keeping them alive. This guide is for the engineer who already shipped a tool-calling loop and now has to answer:

- Why did this tenant cost us $412 yesterday when their plan caps at $40?
- Why is the agent suddenly looping six extra turns on every request?
- Why did the eval pass in CI but the answers are worse in prod?
- How do I prove to security that we are not logging credit card numbers?

GUARDIAN is a 7-phase framework (the acronym intentionally fuses **I**nstrument and **A**nalyze — in practice they ship together):

1. **G — Guard** — circuit breakers, budgets, kill switches
2. **U — Unify** — one observability pipeline
3. **A — Audit** — PII, prompt injection, output validation
4. **R — Recover** — retries, fallbacks, graceful degradation
5. **D — Debug** — trace replay, prompt diffing, deterministic repro
6. **I/A — Instrument & Analyze** — metrics, evals, LLM-as-judge
7. **N — Normalize** — cost attribution, unit economics, runbook hygiene

Every section is built to be lifted directly into your codebase or runbook.

1. The Production Reality Check (with cost math)
2. Phase 1 — Guard: Three-Layer Budget Enforcement
3. Phase 2 — Unify: Picking and Wiring Your Observability Stack
4. Phase 3 — Audit: PII, Injection, and Structured Output
5. Phase 4 — Recover: Retry Taxonomy and Durable Execution
6. Phase 5 — Debug: The Production Debugging Playbook
7. Phase 6 — Instrument & Analyze: Evals with Ragas and LLM-as-Judge
8. Phase 7 — Normalize: Cost Attribution and Unit Economics
9. Alert Pipeline Templates (Prometheus, Datadog, Sentry, Slack)
10. The Single-Page On-Call Runbook
11. Anti-Patterns You Will Regret
12. Appendix: Worksheets, Formulas, Checklists

An agent in production has four failure modes a one-shot LLM call does not: