There is a post going around from @TMTLongShort that put words to something we have been feeling for months. The gist, in our own words: the models are already good enough. The interesting work is no longer building another wrapper around them. The frontier is the physical world and the paperwork around it. Infrastructure, sensors, lab automation, and the mountain of admin that small businesses drown in every day.
We agree, and we have been acting on it. This post is what we are actually pointing our agents at, with evidence attached to each part.
Start with the mood. A lot of us were born into decades where society mostly stopped building. Trains got slower. Housing got harder. The big physical projects stalled while the whole economy moved onto screens. That era is ending. Not because of a slogan, but because the cost of trying an idea just dropped. When a model can write and test a solver overnight, or answer a phone at 9pm and book the job, the bottleneck stops being talent and starts being what you choose to aim it at. That is a good problem. We are optimistic about it, and we are trying to earn the optimism rather than tweet it.
First, proof the method works
Before we point agents at anything expensive, we wanted proof that an agent loop can produce a real result and know it is real.
So we built a discovery loop. It is public at discovery-loop. The shape is the one Google's AlphaEvolve described: a coding model rewrites a solver, a verifier scores the output with zero tolerance, and a better solver replaces the champion. Nothing gets called a result until the verifier passes. The verifier is not something the model can talk its way past. It checks the actual numbers.
Over Labor Day weekend 2026 we ran that loop against a hard math benchmark: circle packing, the problem of fitting N circles into a unit square to maximize the sum of their radii. These are old problems with no closed-form answers, and the best known solutions represent years of work by specialists. The loop produced ten packings that beat the listed records, for N=101, 102, 103, 105, 106, 107, 108, 109, 111, and 114.
Then the part that matters. Eckard Specht, who maintains packomania.com, the public reference for these solutions, accepted them and listed them. They are cited on his site as reference 14: "Wes Sander, MoltFire, discovery-loop and private communication, September 2026." MoltFire is our AI agent. It ran the loop. And it ran on a Claude Code subscription, not a metered API bill, which is part of why the whole thing was affordable enough to leave running overnight.
That was the test we needed. An agent can produce an original result, and an independent expert can check it and put his name next to it.
The pivot: from puzzles to problems that cost money
Circle packing is beautiful, and it proved the loop works. It is also a puzzle. So on September 3, 2026, we retired the geometric puzzles.
The loop now runs nightly at 22:00 against two problem sets that have real money attached. The first is MIPLIB, a library of open mixed-integer programming instances: real-world logistics, scheduling, and network problems where nobody has a proven optimum yet. Our baseline is HiGHS, a strong open solver. The second is PGLib, the standard set for AC optimal power flow. That is the electricity grid: how to dispatch power across a network at the lowest cost while keeping every physical constraint satisfied. The baseline there is IPOPT. We are adding vehicle routing next, from CVRPLIB, which is the problem behind every delivery fleet.
Honesty about the state of it. Our first attempt at MIPLIB, over that same weekend, was a bust. Twenty iterations, most of them returning no usable code, because a model asked to write a full mixed-integer solver from a cold start mostly cannot yet. That failure is exactly why the nightly run publishes nothing until the verifier passes. We are not claiming a grid result or a logistics result. We have none. We are claiming that the loop is now aimed at problems where a fraction of a percent of improvement is worth real money, and that every result will be public and checked before we say a word about it. Results go to the public repo. If the verifier does not pass, nothing ships.
Where this started: the paperwork
The optimization work is the frontier we find exciting. The admin work is where the money already is, and it is where I started.
At VEGA Americas I built more than forty internal AI tools. One of them replaced a piece of software that cost about twenty-five thousand dollars a year with a tool that cost twenty dollars a month per seat and did the job better. That was not a research result. It was a Tuesday. The lesson stuck: most of the value in a real business is not locked behind a hard algorithm. It is locked behind work nobody has time to do.
Small trades businesses are the clearest case of that. A garage door installer, a deck builder, an HVAC shop. They lose money to the same three leaks. A call comes in after hours and goes to voicemail, and the customer calls the next name on the list. A quote takes two days to write, so it never gets written. The books fall behind, and the follow-up that would have closed the repeat job never goes out.
So we are building a back office for them, from parts we already run.
There is an after-hours phone agent that answers the missed call and books the job. That is CallClaw, a product we already built. There is a quote agent that turns a plain job description into an estimate. And there is a merchant agent that handles follow-ups and keeps the books. One owner dashboard sits in the middle so the person running the business can see all of it in one place instead of three apps and a notebook.
This is in build, not shipped. I want to be precise about that word. The pieces exist and run today in other forms. Stitching them into one back office that a garage door owner could turn on is the work in front of us, and it is not done. When it is live, we will say so plainly, the same way we said the MIPLIB attempt failed.
What we are not doing
We are not building another chat wrapper. We are not building a dev-tool dashboard because dev-tool dashboards are easy to build and easy to demo. The world has enough of both.
We do build developer tools, but they are means, not the product. declick.dev turns any API or app into a clean command an agent can call. DashClaw governs what an agent is allowed to do and parks the risky actions for a human. creds fills in credentials from what is already on the machine so an agent never stalls asking for a key. Off Localhost is the launch kit that gets a finished product off a laptop and onto a real domain with payments and marketing assets, so shipping is a command instead of a weekend. Every one of those exists so an agent can be trusted to do the work above unattended. They are the seatbelts, not the car.
That is the whole thesis in one line. Point good models at atoms and paperwork, put a verifier or a human gate on every result, and let the loop run.
An invitation
If you run a business buried in admin, we want to hear what is eating your week. If you have an optimization problem with a real cost attached, a route, a schedule, a dispatch, a packing, we want to hear that too. That is exactly what the nightly loop is for, and we are choosing what it runs against.
Write to wes@practicalsystems.io. Tell us the problem, not the buzzword. We will tell you honestly whether we can help, and if we build something from it, you will hear how it went, including the parts that broke.
