Evaluate agent behavior safely

You cannot put an agent in production if you cannot trust what it does. Learn sandboxing, permissions, jailbreak resistance, and the eval methods that catch real failures.

Overview

You cannot put an agent in production if you cannot trust what it does. Learn sandboxing, permissions, jailbreak resistance, and the eval methods that catch real failures. Octo builds this course around your role, your experience, and what you already know, so the version you get isn't the same one a beginner across the hall is reading.

What you'll learn

By the end, you'll be able to do these, not just have read about them.

  • Sandbox agents so a bad action cannot reach production data

  • Build behavior evals that catch regressions before users do

  • Test for jailbreak resistance, prompt injection, and tool misuse

  • Define what "safe enough" means for your specific deployment

Who this is for

  • You're an engineer or PM whose work now includes shipping AI features.

  • You're a curious operator who uses LLMs daily and wants the substance behind the surface.

  • You're an experienced ML or applied-AI practitioner adding a new specialty.

Prerequisites

  • Solid fluency with the fundamentals, you've shipped or studied this seriously.

  • You're looking to push past intermediate, not refresh basics.

Suggested chapters

This is the typical chapter list. Your version is generated against your background and adapts as you go. It may compress, expand, or reorder these.

  1. 01

    Foundations of Agent Evals & Safety

    The mental model and shared vocabulary you'll lean on for the rest of the course.

  2. 02

    Core building blocks

    The handful of moves that show up everywhere, drilled until they feel obvious.

  3. 03

    Working through real examples

    Applied patterns on examples close to the kind of work you actually do.

  4. 04

    Edge cases & failure modes

    Where the simple version breaks, and how to recognize it before it bites you.

  5. 05

    Putting it together

    Combining what you've learned into something end-to-end and defensible.

  6. 06

    Capstone

    A small project tied to your real work that proves you can use the material, not just recall it.

Real-world projects

  • 01Apply agent evals & safety to a small problem from your actual work or studies.
  • 02Produce one written or built artifact you can put on your resume, portfolio, or in a review packet.
  • 03Run a self-graded capstone against an Octo-provided rubric.

Tools & concepts

Real tools and ideas covered. Octo brings them in when they fit your stack.

  • LLM APIs
  • Embeddings
  • Vector databases
  • Prompting patterns
  • Evals
  • Streaming
  • Function calling

Where this leads

  • 01

    Applied AI / ML engineer roles

  • 02

    Stronger AI fluency in your current role

  • 03

    Foundation for advanced AI specialties

Common questions

  • Is this a fixed course, or is it built for me?

    Built for you. The chapter list below is a typical outline. Your actual course is generated against your role, experience, and what you already know, then adapts as you go.

  • How long does it take?

    Most learners finish in 2–6 weeks at a normal pace, depending on the topic. Octo compresses where you're strong and slows down where you're weak.

  • Is there a fixed schedule or cohort?

    No. You start when you start. There's no live session, no calendar, no deadline.

  • Can I ask questions while I'm learning?

    Yes, every module has an AI Sidekick in the margin. Ask for a different example, push back, or get a clarifying analogy without leaving the page.

  • What do I get at the end?

    A verifiable, HMAC-signed certificate with a public verify page. It records the modules passed, scores, and capstone, not just attendance.

  • How much does it cost?

    Octo is in research preview, courses are open. We'll be transparent before pricing changes.

Agent Evals & Safety, built for you by AI · Octo