a8A8gent
Get the Starter Kit
HomeCoursesClaude Skills + Subagents Masterclass
DA
Course by
Deepak A. · carries the pager
Builders · $247

Claude Skills + Subagents Masterclass.

The deep dive on Anthropic's Claude Skills and Subagents. Configure behaviors, parallelize work, ship agentic workflows that actually use what 2026's frontier models can do.

Keyword
Claude Skills Subagents course
Intent
Buy
Audience
Builders, engineers, AI-savvy PMs
3.5 hr video · 5 lessonsRepo + eval set included7-day money-back
Fast answer

You'll build a production support agent in a weekend. It cites its sources, refuses what it doesn't know, and ships behind a kill switch. The eval set catches the hallucination before the customer does. Median graduate auto-resolves 62% of tier-1 tickets. You will write code; this is not no-code.

Curriculum

Five lessons. One shipped agent.

Each lesson is a video plus a runnable repo. By the end you have an agent in production, observability wired, eval set running on every PR.

L0
Pick the workflow shapeFree
12:04 · scoping doc + 3 patterns
The three workflow shapes that work for support. How to pick yours in 30 minutes using the scoping doc. The trap of trying to do everything in one agent.
L1
Wire retrieval with citations
18:33 · hybrid retrieval + chunking strategy
BM25 plus embeddings, with the query rewriting that actually matters. Citation tags every reply, so audit is one click.
L2
Build the refusal eval set
21:09 · the part that pays for the course
Build a 200-case eval set from your real ticket history. The "I don't know" case is half of it. Score every PR before merge.
L3
Roll out behind a kill switch
15:41 · observability + safe rollout
Langfuse traces. A kill switch that doesn't require a deploy. A rollout dial at 1%, 10%, 50%, 100%.
L4
The 2am runbook
18:20 · what to do when it breaks
The five questions to answer in the first 10 minutes of a page. Three real incidents from our production agents and what we did. A copy-paste runbook for your team.
Full module map

Design, deploy, and improve OpenAI AgentKit workflows with tools, knowledge, guardrails, evals, traces, and ChatKit.

  • 01 · AgentKit Architecture
    AgentKit decision brief

    Understand the toolkit and when to use it.

    • Agent Builder
    • Tools and knowledge
    • SDK vs visual workflows
  • 02 · Workflow Design
    Agent Builder map

    Design the agent before connecting tools.

    • Goal and task
    • Logic nodes
    • Handoffs
  • 03 · Knowledge And Tools
    Knowledge and tool matrix

    Ground the agent in useful sources.

    • Vector stores
    • File search
    • Connector/MCP boundaries
  • 04 · Guardrails And Evals
    AgentKit eval pack

    Control quality and behavior.

    • Guardrail rules
    • Datasets
    • Trace grading
  • 05 · Deploy And Improve
    AgentKit launch plan

    Turn the workflow into a usable product or internal app.

    • ChatKit
    • Monitoring
    • Prompt optimizer loop
Included

Four things in the box besides the videos.

Cloneable repo (TS + Python)

Both languages. Same architecture. Same eval suite. Pick your side, ship the same agent.

200-case eval set

Real-world ticket archetypes including the nasty refusals. Add your own in chapter 2.

Scoping doc template

The 1-page doc that kills bad workflows in a day. Used by every graduate before they write a line of code.

2am runbook (PDF)

The first 10 minutes of an incident. Tested in two production fires we've shipped through.

Two graduates shipped this

Founder · 8-person SaaS

"Killed the Zendesk queue I'd been throwing humans at for 18 months. Eval set is the thing I didn't know I was missing."

Result$11,400 / mo saved
Staff eng · fintech

"Shipped tier-1 in 2 weekends. The refusal patterns chapter is worth the full price by itself."

Result62% auto-resolved
My promise

If chapter 1 doesn't change how you think about evals, I'll refund you. Personally.

One email to Deepak. No form. No follow-up. Refunded in 24 hours. The repo permissions are revoked at the same time. That's the deal.

Objections we hear most

"I can build this myself."

Yes. You can. The course is the six months you'd spend learning what the second draft should have been. Pay for the shortcut or take the long way; both are valid.

"I'll wait until models are better."

The model is the easy part. What you'd be waiting on doesn't change: evals, observability, scoping. Those are the same in 12 months and the same as 5 years ago.

"I bought another agents course and it was hype."

Watch L0 free. If it sounds like the other one, don't buy. If it sounds different, buy. If we're wrong, refund.

"My data is too sensitive for off-the-shelf models."

Chapter 3 covers the PII guardrail and the deployment patterns for VPC / Bedrock / Azure private endpoints. Patterns are model-agnostic.

Starter Kit
$29

Customer support agent template. Repo + eval set.

Most popular
Operators
$197

All 11 modules. 6 agents. The full curriculum.

All-Access
$1,497

5 seats. Vertical playbooks. Team bundle.

Commit

Get the course. Ship the agent. Refund if we're wrong.

Why this matters

Most "customer support AI" demos look great in a sales deck. They look different at 2am when a customer is screenshotting your agent confidently inventing a refund policy that doesn't exist.

The hard part isn't the model. The hard part is the system around it: what it's allowed to say, what it has to cite, how it admits ignorance, how you know if it's slowly getting worse. This course is about that system.

If you can ship a support agent that grounds every answer in your real docs, refuses what it doesn't know, and reports its accuracy every day, you've solved 80% of what every customer-facing AI product needs.

Steps to ship in a weekend

The course is built so you can move at one of two paces. Pick the one that matches your week.

Pace A · one weekend

  • Saturday morning: L0 + scoping doc. Decide on the one workflow.
  • Saturday afternoon: L1 + retrieval wired against a copy of your docs.
  • Sunday morning: L2 + L3. Eval set running, kill switch in place.
  • Sunday afternoon: L4. Roll out at 1% behind a flag. Read the traces over coffee Monday.

Pace B · two weeknights a week

  • Week 1: L0, L1.
  • Week 2: L2.
  • Week 3: L3, L4. Production rollout the next weekend.

Mistakes the course will save you from

  • Building a retriever before you build the eval set. You'll be optimizing nothing.
  • Letting the model decide when to refuse. Refusal is a system decision, not a prompt instruction.
  • Shipping without a kill switch. The day you need one is also the day you can't deploy.
  • Treating cost as an afterthought. The first time it surprises you, it's $4k.
  • Trusting a single eval run. Numbers without a confidence interval are vibes.

Course FAQ

Do I need to know Python?

No. Repos in TS and Python. The TS version is the path most operators take.

Which model is used?

Claude Sonnet by default. Swap to OpenAI / Gemini / local in 5 minutes; the code is provider-agnostic.

What's my running cost?

Per-resolution cost is about $0.01 on Sonnet with hybrid retrieval. Course materials cover both budgeting and the eval-driven cost regression catch.

Can I use this on a Zendesk / Intercom backend?

Yes. Both adapters included; example wires for the Zendesk REST API + a webhook flow that doesn't require any in-app install.