How one of Anthony's AI tools quietly does a piece of work, grades its own work with a strict second opinion, and keeps fixing it until it is genuinely good, with no babysitting.
Imagine you ask a smart assistant to write something or build something for you. Normally you get one draft back, and you have no idea if it is actually right. You have to read it carefully, find the holes, send it back, wait again. You are the quality control, every time.
The Jarvis Loop removes you from that grind. It is an assembly line with a built-in inspector. Work gets made, the inspector grades it hard, anything wrong goes back to be remade, and only finished, quality-checked work reaches your desk. You set the standard at the start and you approve the result at the end. Everything in the middle runs on its own.
The name comes from "Jarvis," Anthony's nickname for his AI assistant, and "loop," because the work goes around and around until it passes.
The whole thing runs on two characters. Think of a writer and an editor.
Writes the plan, builds the page, drafts the email, fixes the code. It is the worker. It never grades its own work, the same way a writer should not be the one who decides their draft is perfect.
A separate AI that reads only the finished work and the agreed checklist, then says pass or fail and points to exactly what is wrong. It is the strict editor who has never met the writer.
They do not collaborate. The Maker hands over finished work, the Checker judges it cold. If it fails, the notes go to a brand new Maker (not the original one), and the cycle repeats. That separation is the whole trick, and the next section explains why it matters so much.
If you ask the same AI "did you do a good job?", it will almost always say yes. It is attached to its own work, just like a person re-reading their own essay misses their own typos. A checker that shares the writer's memory is not really checking, it is agreeing with itself.
So the Jarvis Loop spins up a completely fresh checker every single round. This new checker gets only four things:
It never sees how the Maker was thinking. By default you get two checkers, not one, each hunting a different kind of mistake: one checks that it is correct and the facts are real, the other checks that it actually hits the goal and reads well. For important jobs it scales up to three. Two angles catch what a single reviewer would miss, which is the whole reason a lone checker is not enough.
Here is the full lap, start to finish. Steps are numbered the way the tool numbers them (it starts at zero, because step zero is the setup before any work happens).
Before any work starts, the tool writes a short checklist: the one-line goal, 3 to 8 specific things that must be true for it to count as done, and the hard rules (Anthony's preferences, brand, budget, no em-dashes). It also pulls in the real source material (the project card, the documents, notes from Anthony's "second brain") so both the Maker and the Checker are working from facts, not guesses.
This is the one moment you weigh in up front. You glance at the checklist and either say "yes" or tweak it. After that, it runs on its own.
A fresh Maker gets the goal, the checklist, and the facts, then produces a real draft or builds the real thing. If the job is code, it works in a safe isolated copy so nothing breaks, and it has to leave it in a working, runnable state.
A fresh, separate Checker (or a panel for big jobs) reads the work and grades it against the checklist, quoting proof for every pass. For code, the real test runs first: if it does not build or the tests fail, that is an automatic fail before anyone even reads it. The Checker also gives a 1 to 10 quality score and lists exactly what would make it a 10.
If it passes everything and clears the quality bar, it moves on. If not, the specific gaps get handed to a brand new Maker, and it goes back to Step 1 to fix them. This repeats until it passes or hits the cap (see "Where it stops").
The finished work goes to the "Ready to Review" column, never straight to "Done." You get a tight summary: the goal, the verdict, the quality score, the link, and the one thing (if any) that needs your call. You are the final human sign-off.
Here is the upgrade that just went in. Before, the loop only asked "is this correct?" A boring-but-accurate answer could pass. That is a low bar.
Now the Checker also gives a 1 to 10 quality score, and the work has to clear a minimum bar to count as done: 8 out of 10 for normal work, 9 for high-stakes work (anything going to a customer, anything about money or policy, or anything Anthony flags). The Checker also writes down exactly what is missing to reach a 10, and those notes get fed back to the next Maker.
The tool also keeps a running log of the score across rounds, like "round 1 was a 6, round 2 was an 8." Over time that shows whether the loop is genuinely improving the work or just spinning.
An endless loop is dangerous and expensive, so there are hard limits baked in from the start:
The simple test: if you can list what a great result must include, it can be looped. If "good" is only a gut feeling, turn that feeling into a checklist first, then loop it. A loop with nothing to check against is just an AI agreeing with itself.
A few settings control how strict and how thorough a run is. You rarely touch them, but here is what they mean in plain terms.
| Dial | Default | What it does |
|---|---|---|
| Tries cap | 2 | How many times the Maker gets to fix things before it stops. |
| Quality bar | 8 (9 for serious) | The 1 to 10 score the work must hit to count as done. |
| Panel size | 2, or 3 for serious | How many separate checkers grade it, each from a different angle. |
| Streak | 2 (3 for serious) | How many clean passes in a row it needs when the data can shift. |
| Spend ceiling | capped | The hard limit on AI usage for one run, so it can never run away. |
"Serious" (the tool calls it high-stakes) automatically kicks in for anything customer-facing, anything involving money or policy, or anything Anthony flags. It raises the bar and adds more checkers.
These are its neighbors, not its ingredients. One lap of the loop is only ever two things: a Maker and a Checker. The Jarvis Loop does not run all four of these. The others sit before it, after it, or are baked into the checker. Here is exactly where each one lives:
Big picture: this is Anthony's testbed for "self-driving" AI work. Once it has proven itself on a handful of real projects, the plan is to turn it into a simpler team version so anyone can point an AI at a task and trust the result.