Early access is open. Here's what happens next
Testomatica is opening to a small group of early teams. What onboarding looks like, who it's for, and what's still rough.
The Testomatica blog
Notes on testing at AI speed
- Posts
- 45
- Topics
- 10
- Holiday specials
- 17
Latest
October 2026
Halloween: the scariest bugs make no noise
A crash at least screams. The bugs that keep us up at night return 200 OK and look perfectly fine.
Giving coding agents a QA teammate over MCP
Coding agents can now ask Testomatica what a change will affect, which rules apply and which tests to run, before they write the code.
Measuring 360° coverage instead of guessing it
"What did we test?" deserves a real answer. Here's how a coverage map turns a vague feeling into a number.
Every proven bug becomes a permanent test
Agents explore, a referee proves, and the finding becomes regression coverage. Your safety net grows with every release.
Field notes: our first dry run against a live staging site
Two dry runs, one laptop, one VPN. What the referee confirmed, what the agent got wrong, and what we're changing.
September 2026
An impact graph for code that agents write
A coding agent changes a SQL procedure. Which pages, rules and tests does that touch? We built a graph to answer that from the diff.
Talk Like a Pirate Day: X marks the bug
It's Talk Like a Pirate Day, so here's a bug report written like a treasure map. Turns out pirates had decent instincts about reproducible steps.
Programmers' Day: the 256th day and a second pair of eyes
Programmers' Day falls on day 256, two to the eighth. A small celebration of the people who write code, and of the habit of asking someone else to look.
Why we built on Playwright, and why you can leave us anytime
Every test Testomatica writes is a plain Playwright test in TypeScript. You can run it without us, keep it if you leave, and read it like any other code.
August 2026
The release gate: when a test system is allowed to say stop
A production deploy asks one question: go, stop, or ask a human? Here's when we think a test system has earned the right to answer stop.
A confidence score you can calculate
We don't ask the model how sure it is. Our confidence score is three numbers multiplied together, and you can check the math yourself.
Evidence on every run, even the green ones
A failing run without evidence is an argument. A passing run without evidence is a guess. Here's what we save on every attempt, and for how long.
Staging is not fake production
Staging has real-looking data but none of the real traffic. The rules we follow so tests on staging tell the truth instead of telling stories.
July 2026
SysAdmin Day: thank you for keeping staging alive
The last Friday of July is SysAdmin Day. A thank-you to the people who keep staging running, and a promise about how our tests will treat it.
Budgets, not hopes: keeping agent spend predictable
Agent exploration costs money every step. We cap it per run and per day, and when the money's gone, agents stop and the deterministic tests keep going.
Read the diff before you read the page
An agent exploring a whole store at random wastes most of its time. Ours starts with the diff and the ticket, then decides where to look.
Seven business rules every online store should check on every deploy
Seven checks that hold for almost any online store, why each one matters, and why every rule has to catch a broken page before it can block a release.
Declaring independence from the manual regression checklist
The manual regression checklist did its job for years. On the Fourth of July, a few thoughts on what should replace it, and what shouldn't.
June 2026
An infrastructure failure is not a test failure
When the VPN drops or the model provider is down, nothing is wrong with your product. Our design keeps those two kinds of red very far apart.
Father's Day: fix it once, fix it for good
A dripping kitchen faucet, a patient dad, and the testing habit we keep coming back to: when you fix something, make sure it stays fixed.
First sketches: a browser, a model and a loop
Our first rough prototype of an exploring agent is small: a handful of browser tools, a model, and a log a referee can replay without AI.
Why our referee never uses AI
The part of Testomatica that decides whether a finding is real has no model in it. That's on purpose, and here's why.
May 2026
Oracles, ranked: how we decide what counts as a bug
Not every reason to call something broken is equally trustworthy. Here's the order we trust them in, with one store example for each.
What automated accessibility scans can't see
What axe-core and WCAG 2.2 rules catch reliably, and the parts that still need a person with a keyboard and a screen reader.
Quarantine first: how a new test earns its place in the gate
A new test doesn't block a release on day one. It has to prove it's stable first, and "ten green runs" isn't enough proof.
Mother's Day: "did you check it twice?"
A small Mother's Day story about a stove, a front door, and the oldest verification habit any of us learned.
May Day: the QA work we want to automate, and the part we don't
May Day is about work and the people who do it. A good moment to be clear about which parts of QA we want machines to take, and which belong to people.
April 2026
Earth Day: the greenest test is the one that never calls a model
Model calls cost money and energy. Most testing doesn't need one. A short note on why our regression runs with no LLM at all.
The coverage universe: mapping everything worth testing
Code coverage tells you which lines ran. It can't tell you whether anyone checked the checkout on a slow phone. So we're drawing a bigger map.
Session-based test management, but for agents
Charters, time boxes and debriefs were invented to make human exploration accountable. They turn out to be exactly what an AI explorer needs too.
April Fools: bugs that look like jokes until they aren't
Some bugs are so absurd they read like a prank. That's exactly why they slip through. A short tour of the funny ones and why we check for them.
March 2026
Risk scores are for sorting, not for budgeting
Likelihood times impact is a fine way to decide what to test first. It's a bad way to decide what to skip. Here's how we use it.
Metamorphic testing, explained with a shopping cart
You often can't say what the right answer is. You can say how two answers must relate. That idea is behind a whole family of our checks.
Pi Day math: why we replay three times and accept two
Pi starts with a 3, and so does our referee. A short, friendly walk through the math behind accepting a finding at 2 out of 3 replays.
The testers who taught us how to explore
For International Women's Day, a thank-you to three testers whose books shaped how we think about exploration and whole-team quality.
February 2026
Personas are browser settings, not role-play
Telling a model to pretend it's on a slow phone doesn't make the network slow. Why our personas are configuration first and character second.
Why the LLM should suggest, and code should decide
AI bug verdicts are right about one time in four. Here's the setup that keeps AI useful without flooding you with false alarms.
A love letter to deterministic tests
Everyone's excited about AI agents. On Valentine's Day, we'd like to say a few kind words about the boring tests that actually hold things together.
Checking vs. testing: the distinction that shaped our design
Some work in QA is checking, some is testing, and mixing them up is how AI tools end up crying wolf. How the distinction shaped our design.
January 2026
Data Privacy Day: what our test agents are never allowed to see
A testing agent that can see everything is a liability. These are the privacy rules we wrote down before writing any code.
Blue Monday, red builds: the real cost of flaky tests
A one percent flake rate sounds harmless. Across 150 tests, it blocks most deploys. Here's the math, and what one retry does to it.
What we learned reading 40 papers on AI testing
Two weeks, about forty papers and a lot of coffee. Here's what changed our minds about AI testing, and what didn't.
The verification gap nobody budgets for
Teams pay for coding agents and CI minutes. Almost nobody has a line item for checking what those agents produce.
New year, same bugs: why we're building Testomatica
Code got faster. Checking it didn't. A new year, a small team, and the question we couldn't stop asking over the holidays.
No posts in this topic yet.