Skip to how we do it
Armakuni Contact us

Verify and QA your AI agents and AI code before it counts.

Machines test every change your AI writes, and every AI agent on real cases before each release. Our QA engineers try it like a customer would, an engineer approves each change and each action with weight, and AK Run watches your AI agents live, so you keep the speed and your customers stop finding the mistakes first.

Book a free workshop

Part of AI-Native Managed Services

A one-day workshop. AWS funding typically covers the workshop. An AWS Premier Tier partner, one of about 300 among more than 130,000, validated by AWS for agentic AI consulting.

Everything your AI makes is tested, tried and approved before the lock opens, and watched once it’s out. The names are examples.

AI made everyone a writer.
Nobody hired the editor.

Last weekA 60-person firm

changes shipped that nobody read

Example

AI writes faster than anyone can read. A growing business buys seats for AI coding tools and switches on a few AI agents, and the work arrives faster than its one or two senior people can check it. So it goes out unchecked, or it waits for them.

  • A 60-person firm’s operations manager builds a quoting tool with AI over a weekend. On Monday it’s live with the customer price list in it, and nobody has read a line.
  • AI-written code can pass every test it was given and still round a VAT line the wrong way, leave a key in the code or call a package that doesn’t exist. Anyone reading it would have caught them.
  • A booking form can pass every test and still lose the booking when someone presses Back. The first person to try it would have found that.
  • An AI agent can refund an order, pay a supplier or promise a customer a price by email in seconds, and nobody has said yes to any of it. Once it’s live, a model update or a change in your data can change how it answers overnight, and the first to notice is a customer.
  • Or the company bought the seats and nothing ships sooner, because every change waits for the one person who can review it.

The week beside this is an example.

The AI writes it.
We make sure it’s right.

Four checks stand between your AI and your customers. Our engineers run them on AK Run, the platform behind our managed services. Managing your AI agents and cloud keeps them running: uptime, cost, patches and who gets woken at night. This checks that what your AI makes and does is right, before it counts and for as long as it runs. Routine work goes straight through, and anything with weight waits for a person.

Tested

QA, by machines

Your tests, a security scan and your rules run on every change. Evals test your AI agents’ answers on real cases, again whenever a prompt, a model or a tool changes.

It catches a key left in the code, or a model update that changes the answers.

Tried

QA, by people

Our QA engineers use what the AI built the way your customer would, on a test copy and on a phone, before it goes live.

It catches the form that loses a booking when someone presses Back.

Approved

Verification

An engineer reads and approves every change that passes, and every action with weight: a refund, a payment, a price in an email.

It catches the change that passed its tests and is still wrong.

Watched

Monitoring

Once it’s live, AK Run traces every step your AI agents take and scores a sample of their answers, and drift goes to our engineer.

It catches the AI agent that started answering differently on Tuesday.

Six checks from one week, at six kinds of business

An accounting firmVerification

A change that passed its tests

Invoice roundingWritten with an AI coding tool, Tuesday

31of 31 tests passed
  • Our engineerZero-rated lines rounded the wrong way. Sent back, with a new test.
  • WednesdayFixed by the AI, read again, shipped.

Written by AI, read by our engineer

Example

A property managerVerification

A payment over its limit

Maintenance AI agentPay a contractor’s invoice

$2,400over its $500 limit, so it waited
  • CheckedThe job card says done, with photos. The invoice matches the quote.
  • Yes, 9:14For this invoice, this contractor, today.

Asked by the AI agent, approved by our engineer

Example

A wholesalerQA, by machines

A model update, held back

Sales AI agentA newer model, tested before it went live

14of 200 real quote requests forgot the 20-case minimum
  • The evalsRerun on the new model, as on every change. 186 passed.
  • ThursdayIts instructions fixed, 200 of 200 passed, then live.

Answered by the AI agent, held back by the evals

Example

A dental groupQA, by people

A weekend build, tried like a patient

Booking toolBuilt with AI by the practice manager

1thing no test had tried
  • FoundCancelling one appointment cancelled the whole family’s.
  • Monday 9:00Fixed, tried again, live, and now a test.

Built by your manager with AI, tried by our QA engineer

Example

An online shopMonitoring

Answers, watched live

Support AI agentTuesday’s sample of its answers

9of 50 gave last week’s delivery times
  • AK Run, 11:20Flagged the drop. Our engineer found the courier’s new times missing from its sources.
  • By 15:00Added, 50 of 50 again, and those customers sent the right times.

Flagged by AK Run’s AI agent, fixed by our engineer

Example

A recruitment agencyQA, by machines

A change too big to read

CV readerRewritten by an AI coding tool

2,840lines in one change
  • Sent backSplit into 6 changes, each read in one sitting.

Split by the AI, read by our engineer

Example

More of what the checks catch, as examples:

  • A discount code that worked twicea gift shop, tried by a person
  • Staff bank details written to the logsa cleaning company, found by the scan
  • Appointment reminders set to send at 3 a.m.a physio clinic, tried by a person
  • A refund sent to a card that had expireda gym, waited for a yes
  • A returns policy quoted from last yeara furniture shop, watched live
  • Two-step sign-in turned off for adminsa law firm, read by our engineer

Check everything, and nothing ships.
Check nothing, and anything does.

Tests alone aren’t enough. In a 2026 study by METR, about half of the AI-written code fixes that passed their tests wouldn’t have been merged by the people who maintain the code. So the answer is a design for where each check happens. Machines do what machines can, people try what a customer would, an engineer approves what’s left, and AK Run keeps watching once it’s live. Each decision below stops a failure before a customer finds it. Our engineers set it up in your pipeline and at your AI agents’ gateway, and run it every day on AK Run, the way AK AI Factory checks everything it builds.

Invoice rounding, change 482Followed through the six

thing a person caught that the tests didn’t

    Example

    1. 1

      Machines test it before anyone looks.

      Your tests, a security scan, a check of every new package and its licence, and your architecture’s rules run on every change, in your own pipeline. Whatever fails goes back to the AI to fix, with the failure attached, before anyone spends a minute on it.

      • Your AI agents are tested too, with evals: real cases from your business, the normal, the awkward and the hostile.
      • The evals run again on every change to a prompt, a model or a tool.
      • Where an AI grades the answers, we check its grades against your people’s.

      WhyA machine catches a key left in the code, a package the AI made up that someone else can register, or a model update that quietly changes your AI agent’s answers.

    2. 2

      Small changes, each with its reason.

      Every change arrives small, with what it’s for and what it touched, linked to the request that asked for it. A change too big to read in one sitting goes back to be split. We agree the size with you.

      WhyNobody waves a 3,000-line change through with “looks fine”.

    3. 3

      A person tries it the way your customer will.

      Our QA engineers use every change a customer will see on a test copy, against what the request asked for: the form, the checkout, the email it sends, on a phone as well as a laptop.

      • They try what tests rarely do: the Back button, a second tab, a name with an apostrophe in it, a slow connection.
      • What they find goes back with the steps to repeat it, and becomes a test, so a machine checks it from then on.

      WhyA change can pass every test and still lose a customer’s booking.

    4. 4

      An engineer reads every change that passes.

      One of our engineers reads it and approves it, or sends it back with a reason. Your repository won’t merge a change without that yes, and the AI that wrote a change can’t approve it. The depth follows the risk, so a change that touches money, customer data, logins or prices gets a second reader.

      • Our engineers and QA engineers check by name. The same few people learn your code and your rules, so they know what normal looks like for you.
      • Our engineer’s yes lets a change merge, and your own pipeline ships it, as it does today. You can keep a final say on any release.
      • AK Evolve, our way of modernizing code with AI agents, works like this too: every pull request passes quality gates, and every production change needs a person’s approval.

      Why“The tests passed” isn’t the same as “it’s right”.

    5. 5

      Anything with weight waits for a yes, for that action alone.

      At the workshop you decide what your AI agents may do and which of those actions carry weight. Most businesses start with money going out, a message or a promise to a customer, a change to a price and anything deleted. Each gets a limit: refunds under $50 go straight through, and over $50 they wait for a person. Routine work never waits.

      • The yes is tied to the exact action: this refund, this amount, this customer, and it runs out. When the AI agent acts, the gateway checks again that nothing has changed.
      • It’s enforced outside the AI agent, at the gateway every tool call goes through, with the rules as policies that deny by default. Nothing the AI agent reads can talk its way past them.
      • Every waiting action shows who’s on the rota and what they’re being asked. You can keep the last say on anything you choose, such as a change of price.
      • A waiting action never becomes a yes on its own. If nobody answers in the time you set, it keeps waiting, or it comes to you. How long each kind may wait, and who’s asked next, is written down with you.

      WhyYour AI agents ask about the actions that matter and nothing else. A yes for a $40 refund can’t send $400, and yesterday’s yes can’t be used today.

    6. 6

      AK Run watches it live, and every catch becomes a test.

      Every step your AI agents take goes on a trace: the prompt, the model, each tool call, what it cost and who approved it, in your own accounts. AK Run scores a sample of their answers against the evals, on a schedule agreed with you. When one starts to drift, after a model update or a change in your data, it goes to our engineer. AK Run’s AI agents take the first look; our engineer decides.

      • Every yes and every no is kept, with who gave it, when and why. What a person catches becomes a new test or a new rule for the gates, so next time a machine catches it.
      • AI agents earn a higher limit on the record. When an AI agent’s actions are approved unchanged week after week, you can raise its limit, and after a serious mistake it drops back.
      • It all runs in your own accounts, through roles you grant and can revoke, with no standing admin access, and the record stays there. Engineers read code and the details of an action, not your customer records, and personal, card and health data are masked as they’re collected.

      WhyA customer isn’t the first to notice, the same mistake isn’t approved every week, and a limit moves on the evidence.

    One day to a plan.
    Two weeks on your last month.

    You’ll see what the checks would have stopped, and a fixed monthly price to check everything from here.

    What we’ll tell you not to do

    • Don’t put a person on everything. An AI agent that asks about everything gets its questions skimmed, and we’ll tell you which actions to let go.
    • Don’t keep an AI agent that still fails its evals after the fixes. We’ll say so, and the job goes back to a person.
    • Don’t pay us to read what your own engineers have time to read. If a senior engineer of yours has the hours, we set up the gates, the evals and the watch on AK Run, and hand the reading to them.
    • Don’t start here if an app you built with AI tools still needs making safe for customers and moving to AWS. That’s a project of its own, before this ongoing check begins.
    • It starts with a one-day workshop, How to check every AI-written change before it ships. In the morning we map where AI writes code and where AI agents act in your business, and how each is tested and reviewed today. In the afternoon we design the checks: the gates, what a person tries, what has weight and what’s watched live. You leave with a findings report, the design and a costed roadmap, yours to act on with us or without us. AWS funding typically covers the workshop.
    • Then two weeks, read-only. We read last month’s changes and your AI agents’ actions, and run the gates and the evals over them. You get the report: what a check would have stopped, the gates you’re missing, the plan and the price.
    • Then we switch it on, one repository and one AI agent at a time: the gates and the evals in your pipeline, a QA engineer on what customers will see, the approval rule in your repository, the limits at the gateway, a named engineer on every change, and AK Run watching it live.
    • Every month you get the record: what was approved, what was sent back, what the QA engineers found, what drifted, what became a test and which limits can move.

    Checks we’ve already
    put in at our clients.

    At a government case management platform, AI does the form work and a person approves what it made. A Medicare enrollment company with 600 people on the phones now ships through governed releases. DigitalStrike cut release coordination from weeks to days.

    A government case management platform

    Case management for agencies and non-profits

    80 to 90%

    Onboarding a new client took six to seven months, configured by hand. Generative AI now does the form and checklist onboarding, and a person edits and approves what it made. 80 to 90% of that effort is automated, and the rest is the human step. Onboarding went from about six months to weeks, largely funded by AWS.

    A Medicare enrollment company

    600 people on the phones, in a regulated setting

    Governed releases

    A previous vendor left frequent bugs and uncontrolled releases. We put in CI/CD pipelines, pull-request workflows and release governance, with the first deployment model in place for protected health information. 600 people now work on a stabilized platform, with governed releases.

    DigitalStrike

    A marketing platform for hundreds of car dealerships

    Weeks to days

    We automated deployment, put the infrastructure into code and hardened tenant isolation with a new security model. Time to market for new features fell 30%. Release coordination went from weeks to days, with zero cross-tenant exposure since the new model went live.

    Checked by AWS. An AWS Premier Tier Services Partner, validated by AWS for agentic AI consulting, with more than 500 customer launches on AWS and 1,300+ engineers.

    How AI-Native Managed Services works Read the case studies

    The automated testing has found bugs before they reach production, which saves us so much time and stress. We’re able to focus on growing the product because we don’t have to worry about DevOps or testing.

    Ramin R., OwnerShoppa’s

    What owners say
    to put it off.

    Got another? Bring it to the workshop.

    Our AI coding tool already reviews its own code.

    An AI reviewer is a useful gate, and it sits with the others. But the AI that wrote a change can’t approve it, and a second AI can’t tell you whether it’s what your customer needed.

    Our engineer reads what passes the gates, whichever AI coding tools your team uses. The check happens where a change lands, in your repository and your pipeline.

    Our engineers can handle it.

    If they have the time, they should, and we’ll say so at the workshop. Today the queue is usually one senior person, and the AI writes faster than one person reads.

    We can set up the gates, the evals and the watch on AK Run and leave the reading with your team. Or we take the reading and the QA, so yours get back to building.

    Won’t all this checking slow us down?

    Less than the queue does today. The machines take the routine load, so a person only reads and tries changes that already pass your tests and your rules.

    Small changes are read in minutes, not put off till Friday. Routine actions never wait. Only the ones with weight do.

    We’re too small to need QA.

    You don’t need a QA team. You need the checks on the changes and the actions that matter, for a fixed monthly price.

    A business with one senior engineer is where a mistake that ships unread costs the most.

    There’s no budget for it this year.

    Start with the workshop. AWS funding typically covers the workshop. The two weeks’ report then shows what the checks would have stopped in your own last month.

    The price is a fixed monthly one, which comes with that report, and it grows only with the repositories and AI agents in the plan.

    Let the AI do the work.
    Let a person say yes.

    Bring us last month’s changes and your AI agents’ actions. In one day with the engineers who’d do the checking, you’ll see where AI writes and acts in your business, which actions carry weight, and what we’d test, try, approve and watch first. AWS funding typically covers the workshop.

    Book a free workshop