Welcome to Reliability and Evaluation, the eighth and final module. You’ve built an agent that loops, uses tools, remembers, plans, retrieves, and coordinates specialists. It works — in a demo. But the gap between “works when I try it” and “works for real users, every time, affordably” is wide, and it’s exactly the gap that separates a flashy prototype from something you’d ship. Closing it is what this module is about.

You’ll add the four things every production agent needs. Guardrails: check what goes into the agent (refuse out-of-scope requests before spending a cent) and what comes out (validate the answer, repair or refuse if it’s bad). Retries and cost control: recover from transient failures with backoff instead of crashing, and cap steps and tokens so a runaway loop can’t burn your budget. Evaluation: build a test set and score the agent — often with an LLM as judge — so you can measure quality and catch regressions instead of hoping. The module, and the course, end with a capstone: a production-ready Atlas that assembles memory, retrieval, multi-agent coordination, guardrails, retries, budgets, and evaluation into one system.

Every reliability pattern here — the guardrail refuse-and-repair flow, exponential-backoff retries, the token budget, and the LLM-as-judge evaluation harness — is real, runnable Python, verified end to end against an SDK-shaped mock. Start with Lesson 1 on why reliability and evaluation are what make an agent real.