Agentic QA for Professionals · 4-week live cohort

AI didn't take your QA job. It's turning into Agentic QA.

Run nine agents against a real store. Build a finance agent on real money rails. Ship an application with the agentic SDLC.

Now enrolling

Enroll NowWatch the intro4:21
Starts November 12, 2026Thursdays & Fridays11:00 AM – 2:00 PM ET
5.0 from 24 student reviews
214seats across three cohorts9agents, each on its own model44MCP tools you drive in class4,000+commits by your instructor this year
Where our graduates work
AppleAmazonMetaVerizonGoogleCapital OneIBMJPMorganPwCAccentureCharles SchwabMicrosoftCiscoDeloitteUnitySAPOracleBank of AmericaWarner BrosT-MobileAppleAmazonMetaVerizonGoogleCapital OneIBMJPMorganPwCAccentureCharles SchwabMicrosoftCiscoDeloitteUnitySAPOracleBank of AmericaWarner BrosT-Mobile
The plan

Four weeks. Three builds. Nothing you cannot prove.

Every build ends in an artifact a script can check and an employer can open. The agent's summary is never the evidence.

Weeks 1–2 · Nov 12–20

Build 1: a QA team made of agents

Nine agents on the Zinc Store, a real checkout with a storefront, your own Stripe sandbox and a CRM ledger. Each agent runs on its own model with only the tools it needs. They research, plan, write cited test cases, judge them, publish them to ZincTM, automate them in Playwright, investigate every red, file the defect with evidence and heal a locator without touching an assertion. You approve what ships.

Proof: check:lab green, cases and defects in ZincTM, a CI gate that goes red on a planted regression.

Week 3 · Dec 3–4

Build 2: the finance agent

An agent with MCP hands on money. It pulls twelve months of transactions from your ZincBank account, categorizes them, finds where the money leaks and writes the statement a bookkeeper would. Then it audits the store: every order against the Stripe payment and the ledger, catching the planted money bugs, a replayed webhook that doubles spend, a refund that never reduces it, a double-click that charges twice. It runs on restricted keys: the auditor can only read, and the refund tool needs a human to confirm. Then you test the agent itself: traces, a failure taxonomy, a judge that agrees with you.

Proof: a reconciliation report that matches all three witnesses, the spending analysis, an eval report, injection attempts blocked.

Week 4 · Dec 10–11

Build 3: ship one with the agentic SDLC

Take a spec for an application of your own: a registration site with a database, an intake form and a confirmation email, or the bookkeeping app that replaces the one you pay for. Decompose it into sessions a machine can run while you sleep: every human decision pushed to the front, every exit criterion a command with an exit code. Let the agents build behind the gates. Point your pipeline at the result. Launch it on demo day with a public URL and a green CI run, and defend it out loud.

Proof: the live URL, the CI run, the spec that produced it, and your defense.

What a class looks like

Receipts, not slides.

Three moments from a real run of the lab you will do in week two. Nothing here was staged for the page.

The agent's attempt to edit pipeline/approvals.json is denied by a permission rule
The harness holdsThe agent tried to flip two rejected cases to approved. The deny rule stopped it.It then asked a person for permission. That is the whole lesson: rules in a settings file, not in a prompt.
ZincTM defect ZTM-25: a $100.00 cart is charged $113.99 after checkout showed $106.00
The money bugA $100.00 cart charged $113.99 after the checkout showed $106.00.Filed in ZincTM by the defect-reporter agent with the rule it breaks, the trace, the screenshot and the spec. A developer can fix it from this page.
failures.md: TC-07 expected a discount of −$10.00 and received −$9.99
One centExpected −$10.00. Received −$9.99.The client floors the coupon discount, so what the shopper sees is not what is charged. A read-only judge decided the product was wrong, not the test.
Proud member of the
Claude Partner Network

Claude Code is one of the agents you operate in the cohort. We build and teach with it as an official partner.

The honest part

AI already types the tests. The paid work is the harness.

Someone decides what the agents may touch, what they must cite, how much they may spend, and what proves they did the job. That someone is the one QA role that grows because of AI. It is also the one most teams have not hired yet.

What AI already does
80%

The typing. Agents draft the cases, wire the selectors, run the suite and write the first report.

What you are paid for
20%

The harness. Rules, gates, cited evidence and a budget. You decide what correct means.

The architecture

Nine agents.
One human who decides.

This is the team you run in week two, as it exists in the lab today. Each agent has its own model, its own context and a tool list the runtime enforces. The two judges cannot touch what they judge.

Context researcher

Gemini Flash-Lite

Reads the requirements, the team rules and the student's ZincTM workspace, and writes the report every later agent starts from.

Can changeNothing. Read-only, 12 tool calls max.

Test plan creator

Kimi K2.7

A traceability table, one row per rule, and the questions the team must answer before anyone writes a case.

Can changeOne file: pipeline/test-plan.md

Test case creator

GLM 5.2

Eight to twelve cases. Every case cites a rule by id and states an exact expected value. Amounts computed in cents.

Can changeOne file: pipeline/cases.json

Case evaluator

Claude Haiku 4.5

A fresh context on a different model. It owes the author nothing. It rejects anything uncited, unverifiable or wrong.

Can changeOne file: pipeline/evaluation.json

Publisher

GPT-OSS 120B

Publishes only the cases a person approved in the review gate. If approvals.json is missing, it publishes nothing.

Can changeOne tool: tm_create_test_case

Automation engineer

Claude Haiku 4.5

Turns approved cases into Playwright specs. It asks the live store for test data and never reads the store's source.

Can changeSpecs in tests/lab, linked to ZincTM

Failure investigator

GLM 5.2

For every red test it decides: product or test? It recomputes the expected value from the rule before it says so.

Can changeNothing. Read-only by the runtime.

Defect reporter

Gemini Flash-Lite

Files the product bugs in ZincTM with the rule, the steps, the trace and the screenshot. Never a bug nobody found.

Can changeOne tool: tm_file_defect

Test healer

Qwen3 235B

Repairs a broken locator and proves it with a run. A script checks it did not weaken, skip or retry a test.

Can changeLocators only, never an assertion

The gates

Guilty until proven innocent. Four scripts read what the machines wrote, never what the agent said.

  • npm run rtmTraceability by script. Every rule mapped to a case, or flagged. Not an opinion.
  • npm run reviewThe human gate. You approve or reject each case with a reason. Agents cannot edit the file.
  • npm run heal:checkNo assertion weakened, no wait added, no test skipped. Values only from the approved case.
  • npm run check:labProof from the recordings, the receipts and ZincTM. It never reads the agent's summary.
The curriculum

Eight live sessions. Each ends in proof.

Thursdays and Fridays, 11:00 AM to 2:00 PM ET. No class Thanksgiving week. Every session recorded.

Nov 12
01Proof beats prediction
The blind run: the agent invents test ids and the contract proves it. Your first purchase passes three witnesses: the store page, Stripe and the ledger. Your first harness rules.
Witness receipts for three orders; a table of what the agent claimed against what the systems recorded.
Nov 13
02Build the harness
Deny lists and the gate that blocks a sleep. Context poisoning in three acts. One skill measured with and without it. The cost of every task, and the cheapest model that can do it.
A one-page AGENTS.md you wrote, a skill with a before-and-after score, an activity inventory.
Nov 19
03Cited cases and the judge
Context researcher, planner, case writer, then a fresh-context evaluator on a different model. You plant an uncited case and watch it get rejected. Traceability by script. Publish to ZincTM through MCP.
Cited cases in ZincTM, the traceability output, the evaluator's rejection log.
Nov 20
04Automation and the planted defect
Automation engineer, read-only investigator, defect reporter, healer. The first red is your own test's fault; the second is a planted money bug. A CI gate that goes red on a seeded regression.
A generated suite, a healer audit, a defect filed with evidence, a red-then-green CI run.
Dec 3
05The finance agent, part one
MCP hands on money: your ZincBank account, the store's Stripe sandbox, the ledger. Pull twelve months of transactions and find where the money leaks. Restricted keys: the auditor reads, the refund tool confirms with a human. Reconcile every order. Catch the replayed webhook, the refund that never lands, the double charge.
A spending analysis and a reconciliation report that matches all three witnesses; the key policy you wrote.
Dec 4
06The finance agent, part two
Now test the agent itself. Read its traces. Build a failure taxonomy. Label a set, write one judge per failure mode, and measure whether the judge agrees with you. Try to inject it through a poisoned ticket.
An eval report; injection blocked or flagged; one failure explained without AI.
Dec 10
07The agentic SDLC
A spec with every human decision at the front. Session plans a machine can run. A plan lint that refuses conditional work. Kick off the overnight build behind gates, then point your pipeline at what it built.
The spec, the session plans, the lint output, the build log.
Dec 11
08Launch and defense
Deploy with a public URL and a green CI run. Demo day in pods. An oral defense where the agent's output is changed live and you explain what the gates caught. The credential links to all of it.
The live URL, the CI run, the defense rubric, your metrics ledger written as résumé lines.
Heard in class

From the October tape.

Lines from the current cohort's live sessions, October 2026, as spoken.

Writing code is cheap. Making sure code works is more expensive. And that's what we're learning right now.

Kuzzat Altay · Oct 9, 2026

Asking is like 1% of the problem. 99% is making it safe.

Kuzzat Altay · Oct 16, 2026

Show me the proof. Show me the log. Show me the trace. Show me the execution log. Not your opinion.

Kuzzat Altay · Oct 15, 2026

That topic of agent harnessing is the only thing that makes you unique in the job market.

Kuzzat Altay · Oct 15, 2026

You cannot orchestrate broken agents.

Kuzzat Altay · Oct 15, 2026

Your GitHub is going to be your new LinkedIn. No résumé is going to be better than that.

Kuzzat Altay · Oct 8, 2026
The interview

These are the questions. Can you answer them yet?

Eight of the 386 in the cohort's interview bank. Every answer traces back to a moment in a real class.

  • An agent reports that all tests pass. How do you verify that claim?
  • How would you review a thousand generated test cases when you cannot read them all?
  • How would you design a definition of done that an agent cannot fake?
  • Your company will not send data to a third-party API. What are your options?
  • Why are some models dramatically cheaper than others when quality is close?
  • The agent can write the framework. Who decides the architecture, and why does it matter?
  • How do you verify a UI action really worked, rather than just looked like it worked?
  • How would you detect a hallucination automatically?

There are 378 more. You leave with all of them.

Who's teaching this

One teacher. The one who built the stack.

I'm Kuzzat Altay. I founded CYDEO and trained more than 14,000 engineers across 36 countries. Everything you touch in this cohort, I built and run: ZincBank, ZincTM and its MCP server, the Zinc Store lab, the AI Lab, the CRM that runs our business.

This year I have pushed more than 4,000 commits, and most of them are not QA. I built the LMS our students learn on in eight weeks with a coding agent. I run my company's finance, legal drafting and marketing on agents I harnessed myself, under the same rules you will learn here. Every mistake in this course, I made first.

I teach every session live, and I read every résumé myself.

What we teach here, we run in production.

GitHubgithub.com/kuzzat-altay
Where I've built and studiedAccentureCGIInovalonHarvard Business SchoolCYDEO Agentic Ecosystem
Kuzzat Altay, founder of CYDEO
Kuzzat Altay · founder
What we ship

Every system in this cohort is ours, and live.

You do not practice on a demo. The bank, the test-management system, the store and the agent runtime are products we designed, shipped and run.

ZincBank
Fintech · the bank

ZincBank

A production-grade bank. Your finance agent reads its ledger through a 12-tool MCP server.

● Live · zincbank.cydeo.io
ZincTM
Test management · 29 MCP tools

ZincTM

Requirements, cases, runs, defects and coverage. Your agents publish into your own workspace.

● Live · zinctm.cydeo.io
Zinc Store lab
The lab · your Codespace

Zinc Store lab

A real checkout with planted bugs, your own Stripe sandbox and the nine agents, in a Codespace you get on day one.

● Live · in class
CYDEO AI Lab
AI · developer tools

CYDEO AI Lab

The coding agent and model gateway your credit runs on.

● Live · ai.cydeo.com
Nuebase
CRM · SaaS

Nuebase

The CRM that runs our whole cohort business.

● Live · nuebase.com
Noorvy
Family tech · in build

Noorvy

A first phone for kids 10 to 13, minus the harm: a real Samsung Galaxy, run by parents from their own phone. Built with the same agents and gates.

● In build · noorvy.com
Be honest with yourself

This isn't for everyone.

Come if

  • You're a manual or automation tester ready to add the AI layer.
  • You keep hearing “agentic,” “MCP,” “harness” in interviews and want to actually operate them.
  • You're comfortable with Java/Selenium or Playwright. You already have the 80%.
  • You want a finance agent and a shipped application in your own GitHub by December.

Skip it if

  • ✕You want a six-figure job handed to you for little effort.
  • ✕You expect the full six-month career path compressed into four weeks.
  • ✕You won't put in three to four hours between the live sessions.
What you keep

Four weeks of class. Three builds in your GitHub.

The sessions end. This is what a seat leaves you with.

Code

Your harnessed pipeline

Nine agent files, the deny list, the gates and the scripts, in a repository you own.

Code

Your finance agent

MCP tools over Stripe, the ledger and ZincBank, the restricted-key policy and the eval suite that grades it.

Launch

A shipped application

A public URL, a CI run and the spec that produced it. Built with the agentic SDLC, tested by your pipeline.

Evidence

Receipts an employer can check

Witness receipts, ZincTM cases and defects, agent traces. A portfolio of proof, not claims.

Interview

386 interview questions

Every answer traced to a moment in class, plus the 74-topic study guide and a mock-interview platform.

Access

ZincTM and every recording

Your own ZincTM account and its MCP server, and every session recorded and shared the same day.

After the cohort

It does not stop at week four.

  • A two-day bootcamp on building the framework from scratch.
  • Ad-hoc sessions whenever the tooling shifts, announced in Discord.
  • Your résumé read by the founder personally.
  • A check-in at three months to hear where you landed.
Student reviews

In their own words.

24 students have reviewed the course. We publish reviews of every star rating.

5.024 reviews
  • 524
  • 40
  • 30
  • 20
  • 10
Read all 24 reviewsTook the course? Write a review
I was able to help implementing entire agentic qa workflow at my work for a very complicated web application.
Hakima A.AQ-03 · Quality Analyst
Read the full review
CYDEO gives 1000% ROI. I have taken Playwright and Agentic classes. CYDEO gave me a thousand times the value of my investment.
Ali ParmarAQ-03 · Senior SDET · LinkedIn
Read the full review
…we dove deep into the AI capabilities, and hands-on tried out creating an application from scratch, setting guardrails and running agents.
Ekaterina StepanovaAQ-03 · Software Engineer · LinkedIn
Read the full review
What it costs

One flat rate.

Everything included: eight live sessions, the three builds in your own GitHub, and $50 in AI credits to run them.

Working QA professionalEarly-bird
$999

Early-bird tuition is $999 through Sat, Oct 31, 11:59 PM ET. From Sun, Nov 1 it is $1,299.

  • $949 tuition + $50 AI credit
  • 8 live sessions over four weeks, all recorded
  • The pipeline, the finance agent and a shipped app, in your GitHub
  • 386 interview questions, the study guide, your ZincTM account
Enroll Now

30-day money-back guarantee

CYDEO alumni: sign in with your alumni email at checkout and your alumni tuition is applied automatically.

Employer paying? Get an invoice →

Employer Pays

Your training budget can cover this cohort.

Most QA and engineering teams get a yearly training budget. Ask your manager first. Then we handle the paperwork: your company gets an invoice with its legal name, address, and VAT or tax ID.

Ask your manager

01

Send the ready-made email. It covers the course, the cost, and what you will bring back to your team.

Request the invoice

02

At checkout, choose “Invoice my employer” and add your company’s billing details. It takes about two minutes.

Your seat is held

03

Your seat in the next cohort is held right away. Your company pays by card or bank transfer within 15 days.

Request an Invoice
The invoice showsYour company’s legal nameBilling addressVAT or tax IDPO numberW-9 on request
The record

Three cohorts since July. The fourth starts November 12.

3
cohorts run, July to October
214
seats, most of them working QA engineers
386
interview questions written from class
Who we train

We have done this 14,000 times.

Most of them started with no tech background at all. If we can take someone there to a QA job, we can take you from the 80% to the 20%.

Stay-at-home mom → Test Automation Engineer
Tifa

“My daughter started asking where the money comes from. Now she sees her mom working in tech.”

Read her story
Tifa
Military veteran → Software Quality Assurance
Bradly

“The discipline from the military helped, but CYDEO gave me the skills. Completely changed my life.”

Read his story
Bradly
Babysitter & odd jobs → QA Engineer
Husna

“When I saw my first paycheck, I couldn’t believe it. Nobody in my family thought seven months could change everything. It did.”

Read her story
Husna
Personal trainer → Automation Engineer
Tariq

“As a trainer, I could only count reps. Now I build test frameworks. It showed me how much potential I really have within myself.”

Read his story
Tariq
Delivery driver → QA Engineer
Lilia

“From delivering packages to writing test automation. I didn’t think this was possible for someone like me.”

Read her story
Lilia
2 AM shifts, no work-life balance → Cybersecurity Analyst
Yvette

“Everyone had their doubts. A friend called it a scam. I proved them wrong through hard work. Now I protect data from home with a cup of coffee.”

Read her story
Yvette
Restaurant manager → SDET
Omar

“I was tired of standing. Eight months of study taught me what four years of college couldn’t. Three raises in my first year without even asking.”

Read his story
Omar
Nail tech & cab driver → Automation Engineer
Miriam

“I used to drive passengers to the buildings where I work now. I wasn’t smart, but I had desire. Now I’m making six figures and never missing my daughter’s life.”

Read her story
Miriam
Questions

Before you join.

When does it run?

8 live sessions over four weeks, starting Thursday, November 12, 2026. They run Thursdays and Fridays, 11:00 AM to 2:00 PM ET: Nov 12-13, Nov 19-20, Dec 3-4, and Dec 10-11. There is no class Thanksgiving week. Every session is recorded.

Is enrollment open right now?

Yes. You can enroll and pay today, and your seat is confirmed straight away. This is the fourth cohort, starting Thursday, November 12, 2026, and seats are capped at 60. If you would rather not commit yet, the form lower down puts you on the list for the cohort after this one.

Who teaches it?

Kuzzat Altay, CYDEO's founder, teaches every session live. He built every system you use in the cohort: ZincBank, ZincTM and its MCP server, the Zinc Store lab and the AI Lab. There is no second instructor.

What do I need to know already?

It's an upskill for working SDETs. If you're comfortable with Java/Selenium or Playwright and basic test automation, you have the 80%. We build the agentic 20% on top. You need a GitHub account and a browser. The lab runs in a Codespace we built: the store, your Stripe sandbox, the ledger and the agents are already there. Nothing to install.

What is the finance agent?

An agent with MCP tools over the store's orders, your Stripe sandbox, the CRM ledger and your ZincBank account. It reconciles every order against the payment and the books, and catches the money bugs we planted: a replayed webhook that doubles spend, a refund that never reduces it, a double-click that charges twice. You give it restricted keys, so the auditor can only read and the refund tool needs a human to confirm. Then you test the agent itself, with traces and a judge you can trust.

What do we ship at the end?

An application you build with the agentic SDLC: a spec with every human decision at the front, session plans a machine can run, an overnight build behind gates, and your own pipeline testing the result. On December 11 you launch it with a public URL and a green CI run, and defend it out loud.

How much time does it take?

Eight live sessions of three hours, plus three to four hours a week finishing the labs. Every session is recorded, and the labs come with a catch-up command if you fall behind.

What's the $50 token credit?

Every seat includes $50 in AI token credits for the models you run in class. You practice on those credits instead of paying out of pocket.

Which AI models do we use?

Each agent in the pipeline runs on a different model: Claude Haiku, GLM, Gemini Flash, Kimi, Qwen and GPT-OSS, through the CYDEO AI Lab. Claude Code appears from the instructor's seat for hooks and headless runs. The point isn't one brand. You build an activity inventory that maps each task to the cheapest model that can actually do it, which is what makes a manager trust you with the bill.

Why isn't the upskill just free?

Because paid students finish. In our experience, people who pay anything complete at far higher rates than a free cohort, roughly 80% higher for us. With the $50 token credit built in, your real tuition is almost nominal. The small commitment is what carries you to the end.

I'm a CYDEO alumnus. Is there alumni tuition?

Yes. Sign in with your alumni email at checkout and your alumni tuition is applied automatically.

Does it stop after four weeks?

No. There's a two-day bootcamp on building the framework from scratch, and ad-hoc sessions whenever the tooling shifts. Kuzzat also reads résumés personally, and we check in at three months to hear where you landed.

You already have the 80%.

Build the other 20%. Then ship something you can show.

Enroll Now