Skilljar is a learning management system that hosts our educational content. You're logging into it to access the Anthropic course materials. This separate platform allows us to provide interactive learning experiences, track your progress, and ensure you have access to all course resources in an organized way.
Production Engineering, Evals & Security
Learn to take an agent that works in development and prove it holds up under real production traffic.
Agents that work in development don't always survive production. This module covers the gap that can happen when the system meets an input shape no one tested, hits a rate limit at peak, or reads a fetched page carrying an instruction aimed at the agent. It teaches the engineering judgment needed to close that gap through evals, tracing, failure handling, cost control, and security, so a working prototype becomes something you can defend in a review.
Learning objectives
By the end of this module, you will be able to:
- Write an eval suite that defines what "done" means for a Claude feature before you deploy it, pick the grading method that fits the task, and calibrate an LLM-as-judge scoring against human-labeled cases so the result is one you can defend
- Build a test and tracing layer that catches regressions at the unit, functional, integration, and end-to-end levels
- Create an application resilient to production failures by distinguishing retriable errors from terminal ones
- Keep a system inside its cost, latency, and reliability budget, including when work is spread across several coordinating agents, by instrumenting each call and reaching for parallel agents only when the task needs them
- Defend an integration against prompt injection, jailbreaks, untrusted input, scoped identity, exposed secrets, and data boundaries so the deployment survives a security or compliance review