01 · Support
Customer support agent
Tier-1 inbox with grounded answers, citations, and a refusal that doesn't sound like a chatbot.
Outcome62% auto-resolved
a8Production AI agents
Ship one agent that makes money. Then ship the next one.
The opposite of YouTube tutorials. A8gent teaches what happens on day 14 when the agent breaks at 2am: workflow selection, guardrails, evals, observability, rollout. No demos to a kanban board.
One-time paymentLifetime access7-day money-back
Course preview
module 02 / 11Build
Customer support agent that doesn't lie
Retrieval gating, refusal patterns, and an eval harness that catches hallucinations before they reach the user.
- Pick the right workflow shape12:04
- Wire retrieval with citations18:33
- Build the refusal eval setin progress
- Roll out behind a kill switch21:09
- Observe in Langfuse and react15:41
3 of 5 lessons complete
Open moduleBuilt around the tools engineers actually ship with
trace / support_agent / run_8f2a91
user :: where is my order #A-2841?+ 0.00s
tool :: fetch_order(id="A-2841")+ 0.41s
> status=in_transit eta=2026-06-05 carrier=ups
tool :: retrieve_policy(q="late delivery refund")+ 0.63s
> 3 chunks · cite=policy_v4#refunds
guard :: refusal_check passed+ 0.71s
> grounded=true risk=low
reply :: shipped, citations attached+ 0.84s
latency
842 ms
tokens
3.1k
cost
$0.012
eval
9.4 / 10
A real trace, not a screenshot
Day 14, 2am, the agent breaks. This is what you'll know how to read.
Every lesson ends with one of these. Inputs, tool calls, guardrail decisions, latency, cost, eval score. If you can't see what your agent did, you can't ship it.
- Latency, broken out by tool
Find the 300ms you didn't know you were spending on the wrong retriever.
- Guardrail decisions, before the reply
Refusal, grounding, PII, and the four other checks every production agent should run.
- Evals tied to real traces, not synthetic
Build the eval set from yesterday's failures, not from prompts you wrote in advance.
What you'll build
Six agents. Each one shipped, scoped to one workflow, and measured.
No survey of frameworks. You pick the workflow, build it, eval it, and put it behind a kill switch. The metrics under each tile are what graduates have actually moved.
02 · SDR
AI SDR that books meetings
Inbound qualifier plus outbound sequence agent with reply-handling and CRM writeback.
Outcome3.4x meetings booked
03 · Voice
Voice agent with handoff
Inbound voice on Vapi, with barge-in, function calls, and a clean handoff to a human.
Outcome$0.07 / call
04 · MCP
Your own MCP server
Ship a typed, auth'd tool server that any client (Claude, Cursor, your agent) can call safely.
Outcome11 tools live
05 · RAG
Agentic RAG, done right
Hybrid retrieval, query rewriting, eval-driven chunking. No more "we tried RAG and it didn't work".
Outcome+38 pts ndcg@10
06 · Ops
Internal ops automation
An agent that does the spreadsheet your ops lead does every Friday. Including the part where it asks for confirmation.
Outcome6 hrs / wk saved
The system
Pick the workflow. Guard it. Build it. Ship it.
Four steps, in order. Skip one and the agent is a demo. This is the same loop you'll run on every workflow you ever build with us.
1.0
Pick the workflow
Most agents fail because the scope is wrong, not the prompt. A 1-page scoping doc that kills bad ideas before you build them.
2.0
Guardrail it
Refusal, grounding, PII, cost ceilings, retry budgets. Decide what the agent is allowed to do before it does it.
3.0
Build it
Tools, retrieval, orchestration. With an eval set written from real traces from day one, not the day before launch.
4.0
Ship it
Kill switch, observability, rollout %. A runbook for the 2am page. And the cost dashboard your CFO will ask about.
What this isn't
- -A LangChain tour. We use the smallest thing that works.
- -"Prompt engineering" pamphlet. That's chapter 0.
- -A model comparison thread. Use the one that ships.
- -A cohort with homework and live Zooms. Watch on your time.
- -"10x your agency" hype course. We talk about margins.
- -A theory book. There's a repo for every chapter.
Who this is for
Three people read this site. Find yourself.
Operator
Owners & ops leads
You want one workflow live by Friday. Not a stack. You're fine writing some YAML, not fine writing a graph in Python.
For you if
You can describe the workflow but can't get an engineer to ship it for under $40k.
Skip this if
You're allergic to the terminal. We'll lose you in chapter 3.
Builder
Engineers & indie devs
You've built a demo. You're not sure what production looks like. You want evals, observability, security, and a cost story.
For you if
Your demo works on the happy path. The PM wants the rest of the path.
Skip this if
You already run agents at scale with a paged on-call rotation. We're behind you.
Agency
Selling outcomes, not hours
You want vertical playbooks, scoping scripts, and a way to price by outcome. You don't want to retrain your team on a new framework every month.
For you if
You bill hourly and want to bill on retainers and revenue share.
Skip this if
You sell "AI strategy" decks. We don't ship those.
Pricing
One-time payment. Lifetime access. 7-day money-back, no questions.
We don't run launch sales. The price you see is the price. The middle tier is what most engineers and operators pick.
Starter Kit
One workflow, end to end.
$29one-time
- Module 02 in full (Support agent)
- Eval harness + starter dataset
- 1 scoping doc template
- No community access
Most picked
Operators
All 6 agents. Full system.
$197one-time
- All 11 modules, 6 agents shipped
- Every eval set + repo
- Private community + monthly drops
- Scoping scripts (operator + agency)
All-Access
For teams and agencies.
$1,497one-time
- Everything in Operators
- 5 seats, transferable
- Vertical playbooks (4 industries)
- Office hours, twice a month
7-day money-back, no questions, no forms.
If chapter 1 doesn't change how you think about evals, email us. Refunded in 24 hours.
Real outcomes
Anonymous on purpose. Specific by design.
We don't run testimonials with headshots and last names yet. These are real outcomes from real graduates, reduced to the part that matters.
Founder, 8-person SaaSsupport agent
“Killed the Zendesk queue I'd been throwing humans at for 18 months. The eval set is the thing I didn't know I was missing.”
Result$11,400 / mo saved
Solo engineerMCP server
“Built a typed tool server in a weekend. Three clients are paying me for access. The auth chapter is worth the course on its own.”
Result$2.1k MRR in 6 wks
Agency owner, 6 staffvertical playbook
“Stopped selling hours. The scoping script alone cut my pre-sale calls from 3 to 1. We close on outcomes now.”
Result2.7x avg deal size
Staff eng, fintechobservability
“The Langfuse chapter caught a $4k/day cost regression on our internal agent. Paid for the team license in 48 hours.”
Result$120k / yr avoided
Honest aggregate
Across 412 graduates in the last 12 months, the median time-to-first-shipped-agent is 11 days. About 28% don't ship in that window. That number is in here too.
Built by engineers
Not a creator economy course. A field manual from people who ship.
Every chapter is written from a system we've shipped, broken, paged on, and fixed in production. We tell you the version we lost to before we tell you the version that works.
runsproduction agents serving 9M+ msgs / moshipped3 venture-backed companieson-callfor 11 years and countingspent$1.4M+ on inference, learned the bill
What you're really buying
A shortcut to the second draft.
You'll save the six months it takes to learn that your first agent architecture was wrong. The repo is the answer key; the chapters are the why.
- A repo per chapter. Cloneable, runnable.
- Eval sets from real traffic, not synthetic.
- A scoping doc that kills bad workflows in a day.
- A 2am runbook for the page you'll get.
Honest FAQ
The questions you'd ask a friend before you bought this.
I already use Cursor. Why this?
Cursor is for you writing code. This is for the agent the user talks to. Different problem, different stack, different failure modes.
Do I need to know Python?
No. Operators can ship the support agent on n8n. Builders get TypeScript and Python repos for the same workflow.
Is this a subscription?
No. One-time payment, lifetime access. Updates included. If we deprecate a chapter, it's because the new one is better, not because we're charging again.
What if Claude / OpenAI changes the API?
The patterns don't. The chapters are updated within two weeks of a breaking change. There's a changelog. There's no "lifetime access asterisk".
Will this make me money?
No course can answer that. We can show you the workflow, the pricing structure, the scoping script. The shipping is on you.
Do I get a community?
On Operators and above. It's a small Discord, no growth-hacked channels, no networking spam. Engineer-to-engineer questions get answered in hours.
Is there a free chapter?
Three free chapters on the course pages. No email wall, no PDF gate. If you can ship from the free ones, ship.
Why no countdown? No "launch price"?
Because they're lies. The price is the price. If we ever lower it, anyone who paid more gets the difference back in credit. That's the policy.
My team is 8 people. Which tier?
All-Access is 5 seats. Buy two if you want 10. Or email us and we'll cut a flat number, no haggling.
How long does it take?
Starter Kit: a weekend. Operators: two to three weekends, with one shipped agent at the end of each. All-Access: same, plus playbooks you'll use forever.
Didn't find yours? Email us. We answer every one.
Ship one agent
Stop watching tutorials.
Ship the agent.
The $29 Starter Kit gets you one workflow in production this weekend. If chapter 1 doesn't change the way you build, you get it back in 7 days.
Not a launch sale. Not a closing window. Just the price.
Median
11 days
to first agent in production
Refund
7 days
money-back, no forms
Repos
11
cloneable, one per chapter
Access
∞
lifetime, updates included