Nobody Asked My Agents to Add These Checks. Not Even Me.

In interviews, people ask me how I handle AI slop. I tell them. About two minutes in, I watch their eyes glaze over.

That is on me. I should have started with the kitchen. Instead I answer with the machinery: audit dimensions, pre-commit hooks, negative controls. All true. None of it lands.

So here is the version I should give.

What is AI slop, in plain words?

AI slop is work that looks finished and is not: it compiles, it reads well, and it is wrong.

Picture your sink. You cook dinner and leave one pan. Fine. Tomorrow you leave two. By Friday the sink is full, and nothing gets cooked until someone spends an evening scrubbing.

Software works the same way, and programmers call the pile tech debt. Every shortcut is a dish. Every bug nobody caught is a dish. Every page that says something the code no longer does is a dish. One at a time, they cost almost nothing. Together, they stop the kitchen.

AI makes the pile grow faster. An agent can cook all night, and it can leave a pan in the sink with every meal. When the cooking speeds up, the sink fills faster too.

How do you stop the dishes from piling up?

You clean every dish as you use it. Automatically.

My repository is the folder that holds all of Orbyt's code and its history. Saving a change into it is called a commit, and checks run first. Each commit runs the tests related to what changed, plus the checks that are allowed to block. If one of those fails, the commit stops. The full test suite runs before every push to the live site.

Most audit dimensions report what they find instead of blocking. An audit dimension is one kind of failure the system knows how to look for.

How I Work Like a Thousand-Person Company, With Zero AI Slop walks the whole system from the top. The short version: nothing is trusted because it looks right. It is trusted because something tried to prove it wrong and failed.

The checks started early. The first commit landed on February 8, 2026. The first tests followed on February 22, 63 of them. The first pre-commit check ran on March 8. The audit began on March 21, 2026, with 16 dimensions.

Who writes the checks now?

I used to add these checks. Now I mostly don't. Since July 2026, most new checks have been the agents' idea.

The agents are AI sessions in Claude Code and Codex, two AI coding tools. They write the code, and they write the checks that guard it. They work under one standing rule I set. When something slips through, build the check that would have caught it. My shorthand for it: mechanize the specific miss, do not write a lesson about it.

A lesson is a hope. A check is a fact. When an agent finds a mistake, fixing it does not end the job. The job ends when a new check exists that would have caught it.

Above the sessions sits Orbyt Collective, which I call The Machine. It is a team of AI agent officers that run departments on a schedule. I still decide money, public claims and the rules. Full autonomy is the goal. We are not there yet.

The agents make mistakes. Of course they do. My tests once passed while the log was wrong. A cleanup step once undid a developer's work. Both are on the record because they got past a check. Each one became a new check, so the same mistake gets caught before it ships next time.

The fair objection arrived in August. I pointed a second AI, Codex from OpenAI, at the setup for one afternoon to find its holes. It told me the agents can edit the checks. An agent can write a check that passes for the wrong reason. So a check is not trusted until it has been watched failing on the bad version and staying quiet on the good one.

By rule, significant work also gets a cold review before it lands. A cold review is a fresh session of the other AI, one that never saw the work being built. I still own the rules about the rules.

How fast did the checks grow?

On August 17, 2026, I wrote this in Codex Accuses. Claude Convicts.: "Eighty-eight audit dimensions. 19,949 tests."

The main test suite has grown from 17,934 tests on August 17 to 38,041 on October 17, 2026. Counting every suite, browser, iPhone, Collective and console included, the site counts 45,153. The audit has had 100 dimensions since September 24, 2026.

Most new dimensions now start with the agents, one miss at a time. A new dimension has to answer a class of failure that has actually bitten us. A guess does not qualify. 84 Ways to Tell Me I'm Wrong. explains what the dimensions read, and why each one must prove it can stay quiet as well as loud.

Why can one change reach every page safely?

Because most of it reads from one source. I call it the flywheel.

The numbers in this post are the example. One generator runs on every commit and writes the site's stats file, and the same counts land on the About page, the founder page, Building in Public, the Collective pages and the zero AI slop page, with nobody typing them.

One edit, many surfaces. The upside is one source for every count.

Read the same sentence the other way and you get the risk. One bad edit breaks many surfaces at once.

The checks keep the upside safe. The commit regenerates the stats file instead of trusting a typed number. A drift check refuses a copy that does not match a fresh run. Nobody has to remember which pages read the file. The flywheel on the zero AI slop page draws the loop.

What happens when a check is wrong?

A measurement of the real system catches it. The clearest case on my record is a monitor.

A health check watched the service that sends account emails. Its query fetched one page of 200 log rows, with no filter. The window held 9,072. Eighteen of them were real failures. The check found none and reported all clear.

The tell was in its own output. The count it reported equalled the page size. A number that matches the limit of the query is a truncation with a number written on it.

One day after the check shipped, a measurement against the live system exposed the gap. The fix was to filter first and count everything. Its test now runs real log lines through the check, and a full page is reported as cut off, never as clean.

The record lives as failure 54 in the failure corpus, the public list of every failure we have documented. Self-healing, as I use the word, means an error becomes a check and the check stays.

What to do Next

Pick the dish you keep leaving in the sink. One mistake that keeps coming back.

Write the check that would have caught it. Make it fail on the bad version. Make it pass on the good one. Then let it run on every commit. Do it again for the next miss.

I used to run that method by hand. Now the agents run most of it, under the rule I set.

For the full picture, read How I Work Like a Thousand-Person Company, With Zero AI Slop. It puts the flywheel, The Machine and the checks on one page, with the live counts.

If you want an outside read on your own agent setup, Orbyt Consulting starts with The Walk, a fixed-scope AI agent assessment. To start a conversation, use the contact page.

Since July, most new checks are the agents' idea, and every one has to prove it can fail.

Related reading:

• How I Work Like a Thousand-Person Company, With Zero AI Slop shows the flywheel, The Machine and the live counts on one page.
• My Tests Passed. The Log Was Wrong. explains why a response and its log needed one check that compared them.
• Our Cleanup Undid a Developer's Work shows how a guard became a writer, and the file level rule that fixed it.
• Codex Accuses. Claude Convicts. explains why one AI reviews another's work before it lands.
• 84 Ways to Tell Me I'm Wrong. examines what the audit dimensions read, and the dimension that audits the audit.