Claude Certified Developer: Ship a Production Agent

Build a real Claude Agent SDK agent from 56% to 88% success and ship it. Maps to the CCDV-F exam.

New

Created byVivek Amilkanthawar

Last updated 7/2026

English

What you'll learn

Build a working Claude Agent SDK support agent from scratch and measure its real task success rate: 56 percent at baseline, 88 percent by the end
Build an eval harness that grades every run against a fixed golden set instead of hand-grading, and catches real bugs that a clean exit code hides
Design least-privilege tools and a context pack that lifted task success from 56 to 84 percent without changing the underlying model
Add memory that survives across sessions inside a hard token budget, and a coherence judge that catches an agent that looks fine but has amnesia
Apply three cost controls (caching, model routing, a budget ceiling) and measure which ones actually pay for themselves versus which make the bill worse
Ship the agent headless with a security layer that survives 8 real attacks, observability traces, and a real incident response, then reuse the architecture

Course content

5 sections • 13 lectures • 2h 13m total length

Claude Certified Developer: The Prototype Chasm8:40
An agent answers a support ticket instantly, with total confidence, and gets it wrong. Why an agent that sounds right is not the same as an agent you can trust, and a tour of the finished agent this course builds toward.
I Built a Claude Agent in One Evening. It Was Wrong Half the Time11:50
The simplest agent that can plausibly work: one model, one system prompt, one tool. Graded by hand against 25 real support tickets for a baseline: 56 percent task success.
Four Tools, No Model Changes: 56 to 68 on a Claude Agent11:19
Four least-privilege account tools, enforced read-only at the data layer, plus MCP in practice. Task success moves from 56 to 68 percent with no change to the model underneath.

The Same Claude Agent Scored 28% and 56%. Which Number Do You Trust?11:46
Why hand-grading doesn't scale and drifts. Building an eval harness with a fixed golden set and receipts (tools called, tokens spent, real dollar cost), and the racy bug it catches that a green exit code missed.
One Missing File Made My Claude Agent Lie About Privacy8:19
A context pack: four budgeted markdown files covering product facts, tone, escalation policy, and known failure modes. Task success moves from 68 to 84 percent, and cost per ticket drops.
Does Your Claude Agent Actually Remember You, or Just Look Like It Does?12:11
Memory that survives across sessions inside an 800-token compaction budget, and a coherence judge whose first version scored a memory-less agent 5 out of 6 by mistake.
I Built a Subagent for My Claude Agent. It Tied and Cost Double12:12
A narrow "investigator" subagent for digging through logs, isolated in its own context window. A focused A/B test: it tied the single agent 5 out of 5 to 5 out of 5, at 1.9 times the cost.

None of My 3 Cost Controls Beat the Claude Agent Baseline12:22
Caching, model routing, and a hard budget ceiling, added and measured honestly. The bill goes up, not down, and the lecture explains exactly why, lever by lever.
I Attacked My Own Claude Agent 8 Ways. It Wouldn't Break10:35
8 attacks, including prompt injection through the agent's own tools, run against a baseline and a hardened agent: 16 runs, zero compromises. Least privilege, a PreToolUse hook, and human-in-the-loop escalation as the actual controls, not the model's good behavior.

Shipping a Claude Agent Headless: No Terminal, No Me, 3am8:50
Making the self-hosted versus managed decision on purpose, then shipping the agent headless: one pure entrypoint, ticket in, reply and telemetry out. Median ticket time 16.1 seconds, p95 at 40.3 seconds.
A Claude Agent Ticket Took 70 Seconds. Here's Where the Time Went8:29
Turning structured run records into a trace and reading a real 70-second ticket through it. Only 19 of those seconds were tool calls; the other 51 was the model thinking and writing.
My Claude Agent's Security Guard Was Silently Doing Nothing8:46
A real incident, run the way you'd run an on-call shift: alert, trace, root cause, fix, regression. The root cause: a permission gate that never checked a single thing, the entire time it was live.

Requirements

Comfortable reading and writing basic TypeScript
A terminal and Node.js installed
An Anthropic API key (a free-tier key is enough to follow along)
No prior experience building AI agents required

Description

This course builds one real production AI agent, start to finish, using the Claude Agent SDK. You watch a single support agent get built, measured, and shipped, lecture by lecture, with every number pulled from actual run receipts instead of slides.

It starts honest. The first version is a one-evening agent that answers real support tickets and gets 56 percent of them right. From there, each lecture changes exactly one thing: tools, an eval harness, context, memory, a subagent, cost controls, security, deployment, and observability. Then it measures what that change was actually worth.

Most lectures report a win. At least one reports that the fix everyone recommends made the bill worse. Both get shown on screen, with the receipts, because a course that only shows wins is not teaching you how production actually works.

The course also maps to the Claude Certified Developer (CCDV-F) certification blueprint. The agentic loop, tool design, evals, context engineering, memory, subagents, cost, security, and deployment are covered in the order the exam expects, built as working code instead of slides.

What makes this course different:

- Every metric on screen comes from a graded run against a fixed set of support tickets, never typed in by hand

- One companion GitHub repo with a checkpoint branch per lecture, so you can run every version of the agent yourself and check the numbers against your own receipts

- At least one lecture shows a popular best practice failing, with the real numbers, instead of only showing what worked

- A full working eval harness, cost guard, and security layer you can reuse to cold-start a second agent from day one

- Built for developers who already know basic TypeScript. No prior agent-building experience assumed

By the finale the agent holds steady at 88 percent task success, with a working eval harness, a cost guard, a security layer that has been attacked on camera and held, and a full architecture you can reuse for your next agent.

You need TypeScript and a terminal. Everything is built from the public companion repo, branch by branch, so you can run every version of the agent yourself instead of taking my word for any of it.

Who this course is for:

Developers who have played with the Claude Agent SDK or Claude Code and want to know what separates a demo from something you can run in production
Anyone preparing for the Claude Certified Developer (CCDV-F) certification who wants the concepts as working code, not just exam theory
Engineers who own, or will own, an AI agent in production and need a working playbook for evals, cost, security, and incident response

Claude Certified Developer: Ship a Production Agent

What you'll learn

Explore related topics

Course content

Foundations3 lectures • 32min

Measure Everything4 lectures • 44min

Harden It2 lectures • 23min

Run It in Production3 lectures • 26min

The Architecture1 lecture • 8min

Requirements

Description

Who this course is for: