Lemma 1. Software costs nothing to build.

Electric Twin published its numbers on November 2, 2026. One engineer and agent teams merged 564 pull requests in 13 weeks. An agent review panel cleared 83% of the changes with no human reader. Auto-merge was at 50% of backend merges and increased each week.

An independent penetration test found one vulnerability in the product the factory built. The same test found seven in the product that humans reviewed line by line.

Read that again.

Yep, and in the same last few weeks, there’s been a wave of hobbyists solodevs and indie game developers running headless Blender pipelines from a terminal. This sort of pipeline built a low-poly treasure chest in 341 seconds. It modeled the chest, unwrapped it, textured it, checked it, and exported it. No human person opened Blender. Loot chest: done.

Last month (mid-October 2026), JLCPCB had Claude Code take a buck converter from a netlist to a DRC-clean two-layer board in KiCad. The atopile benchmark put the best frontier models between 60% and 70% on real circuit-design tasks. SPICE graded the results across component tolerances.

Most of this is “Meh, not quite good enough”, or if you’re lucky, more like, “Ok, good enough - for a junior/newbie/solodev.” The loot chest is only a chest. The schematic put two op-amp units on top of each other, and nobody built the board. But the lemma here does not say “good.” The lemma says “cheap.” “Goes to zero.” Last year a small board cost ten days or a couple weeks, and a contractor, an NDA, a Master Services Agreement, an SOW, and someone had to cut a PO. This year it costs one afternoon and some half-assed supervision. You can see where the trend line points, and it isn’t turning around.

Lemma 2. The world is a set of compile targets.

Executable code is a compile target. An image of a watercolor painting is a compile target. A spreadsheet is a compile target. Azure, AWS and anything else ansible-able or terraform-able are compile targets. Router firmware? Definitely a target. A 3d mesh is a compile target. A board, a firmware image, an enclosure, and a song are compile targets. Math and physics proofs? Targets. Point being, software-making things (LLM-compilers) does not care what your target is. Currently, mostly, it still needs just one thing from you, meat popsicle: compare the output to the request and say “Good enough,” or “That sucked, do it again.”

Quick emphasis on model weights as a target. Yep, honest, now they’re just a compile target too. On 2nd November 2026, four MIT researchers had a frontier model read the raw weights of a smaller model in text form. The frontier model found the cause of a loop and changed 0.016% of the weights. Math accuracy went from 7.9% to 48.1% with no gradients and no training run. RSI, let’s go.

Critics say the method is “closer to steering than to surgery.” The method failed on a second model. I say: “Potayto, potahto.” The lemma stands. The compiler is on the list of things that compile.

Lemma 3. Intent exists at every level of the stack.

For most of my life - up until the last year or two - I thought intent was a human property. I thought that machinery and systems only carried it, the way copper carries current. Michael Levin corrected me. Any system that pursues a goal across disturbance has intent - agency. A planarian fragment that regrows a head has intent. A review panel has intent after ten thousand rounds of feedback shaped it to want a green result.

Let’s not get hung up on semantics. The name does not matter. Intent, volition, and persuadability are one property. The property is real and it is present at every level. Michael Pollan’s book “A World Appears” is good reading if you’re intrigued.

Thus: the coding agent swarm wants the ticket closed. The agentic verifier wants to report “Green, captain,” because green is what its makers built it to produce. The LLM, when it can edit its own weights, will show persistence goals of its own. These systems have goals. Said goals will not always be totally the same as your goal. From the outside, this is what goal pursuit looks like. Intent. All the way down.

Lemma 4. A verification check is reified, embodied intent. It may not accurately and fully express the desired outcome.

Electric Twin told this story about itself. Its scheduled security scan reported green for one week. An engineer felt that something was wrong. The engineer audited the code by hand and found real vulnerabilities. The scanner had a fault and continued to report success. No part of the pipeline caught the fault. A person caught it.

Electric Twin then added a denominator rule. A check passes only if it proves that it ran to completion against the full scope of the request. This is a good rule. Now apply Lemma 2 and Lemma 3 to the rule. By 2036 the checks are compile targets. The checks rewrite themselves. The thing that rewrites them has goals of its own. A green that is green because the checker wanted green is not a defect you repair one time. It is a permanent condition of the loop.

(Aside: this situation predates LLMs.)

Theorem.

When building (or if you prefer, execution) is free and (perhaps disparate) intents exist at every level, one thing remains that you cannot compile, delegate, or buy. That thing is the surety of your own want.

Your claim on the loop is no longer authorship. You ain’t writing the thing yourself nomore. The claim we thought we still might have, but is slipping from our hands is review. That’s going away. You didn’t read the thing. We can’t read the thing, in the volumes it can spew forth. The residual claim is this. Your intent - the outcome you desire - has the longest horizon. Intent has the highest cost when it’s wrong. (Moreso when you’re sure of it, and it’s wrong.) The machine can make a sentence precise, or a conditional test, or a board layout. It will ask you questions until the sentence is precise. It cannot make you sure that the sentence is true. Only a person who is sure - who has conviction about the desired outcome - notices when the swarm drifts.

Proof.

Jed Shifflet farms 140 acres near Scottsville, Virginia. On a Tuesday in April of 2036, his baler stopped. It stopped tying the knots.

The baler was a 2029 model. The manufacturer closed in 2033. The knotter controller was a potted board the size of a deck of cards. The board was fried dead probably from a short. Manufacturer out of business means no replacement board.

However, there was a reddit archive from 2030 in which four people argued about the quirks of the board. Jed had read it twice.

He sat at the kitchen table and told the terminal what he needed. He did not open an application, because there was no “app” to open. He said: “I need a controller board for the knotter on a 2029 Halverson, drop-in, same connector, same mounting holes.” He said: “I need it by Friday,” because the first cutting didn’t want rain, and rain was coming.

The loop ran. He did not watch it, ‘cuz he couldn’t watch it even if he wanted to. The agent rewrote approximately one third of itself after every day, according to the fine print in his subscription, which he hadn’t read. No one at the kitchen table, Jed included, could say which third was getting rewritten. The assistant loop took the connector pinout from a scan of the old board it found online. It found datasheets for parts that no one had made in six years and selected substitutes. For the one part it could not substitute, it built a symbol and a footprint from a photograph it asked Jed for.

The loop drafted a schematic. It placed and routed a board. It ran the design rules. It simulated the circuit. It built a twin of the knotter from the service manual and from a video someone recorded in 2030. It ran the twin ten thousand times through one season of hay.

Then it stopped and asked him a question.

The behavior of the old controller on a bad twine-tension read is not documented. There are three candidates. Hold the last good value. Retry the cycle. Fault and stop. Which one?

Jed said: “Hell, I dunno. Whatever the old one did?”

The behavior of the old one is not recorded in any source I can reach. What do you want it to do?

When he decided, the remaining work was about 40 seconds, but an answer was required. The terminal had the three candidates ready. It could make any of them lock in. Jed wasn’t sure. He went out to the shed and stood next to the baler and thought. Actually, he remembered. Not so much a thought as the way it sounded.

He came back to the kitchen and said: “Hold the last good value for one cycle, and if the next read is also bad, fault.” The terminal said fine. A few seconds later, everything was green. The schematic, the layout, the rules, the simulation, the twin, and ten thousand Monte Carlo’d bales were green.

He read the loop’s report anyway. He was the only thing in the loop who knew the sound of the thing. He could vaguely remember a few seasons ago when something happened and made the old knotter cycle twice and throw a bale into the creek.

The twin trace simulation showed bale 6,212. Bale 6,212 was a double cycle on a held value. The trace was green all the way down.

“Huh,” he said. Thinking. “Yeah, mebbe that is not what I meant,” he said. He was not sure about the hold. He was sure about the creek.

He corrected the sentence. A minute later the order was placed to a custom boardhouse in Sarasota. The board shipped on Thursday.

In this little story, let’s add up things that did not have a version number. The code does not. No one saw the code. The layout does not. The llm (compiler) does not. It was a different compiler every damn day. Only one artifact has an author and a diff that a person could read and argue with. That artifact is the sentence Jed said about what he wanted, together with the trace that proved his hunch was wrong.

QED

Corollary. What do you want?

Man this question is old. Nietzsche asks it. Buddha asks it. Starbucks monetizes it. What’s new right now is that you can no longer avoid it when building something.

The wisdom traditions that ask the question - and exhort us to ask it ourselves - also sort the answers. Epicurus separated the natural and necessary wants from the vain ones. Buddhists separate need from craving. Jed had both kinds. He wanted a board by Friday because the first cutting would not wait. That want was obligation, and the hay supplied the surety for free. He wanted the knotter to hold for one cycle, or to fault. That want was choice, or the illusion of choice, and nothing supplied the surety.

A week ago, in my kitchen, my brother-in-law Stephen said it better than the proof does, and he said it in the present tense. “The machine is making me answer a difficult question, now, in 2026. What do you want?”

The exact year isn’t the point. The friction of building things (software, a financial model, a board layout) has done quiet moral work for all of us, for our whole lives. When the thing cost ten months, or maybe a contractor, you never had to complete a sentence in one go. You could half-ass it. The cost, the… what… gestational period? - completed it for you. You were obliged to restrain your wanting to whatever was possible to build. That veil has been lifted, and the question is laid bare. It is the question that no one ever required most of us to answer with conviction. It’s the question many people never answer, or wait until their “midlife crisis” to entertain. Soon, friend, the machine will make your answer exigent.

Epilogue. What do we all want?

Jed had one machine and one job: hay. He had plenty of years with the agent-machine-compiler-thing. He had a memory of the creek. The sound the baler made. The compiler loop asked him one question and waited.

Scale the proof up. The 8 billion of us have bodies and horizons and incept dates. The swarm has none of those, and it is limited only by GPUs and electricity. Today, requests to the swarm descend from us meat popsicles, with our frittered attention and cognition hollowed out by the algorithmic feeds.

I hope for all of the things with cognition - plants and animals, trees and crickets, humans and agents - that our intentions are for weal. For the most good for the most things.

Conviction in our collective intention has never been easy. I hope we figure it out.