• How will companies domiciled in the U.S. but employing workers overseas (both Anthropic and Centaur are examples) continue to develop product?

• What happens to the pipelines, the half-finished migrations, the live projects that depended on a model now behind a border?

• Does carving the frontier along national lines actually make anyone safer, or just slower?

So the timing makes this a strange artifact: here follows a field report on a tool many of us can no longer reach, at least for a while.

Well, the Fable hors d’ oeuvre did whet the appetite quite a bit.

Same sensors, different insights

The first test was the one closest to our core. Every week, Chiron — the natural-language layer of our Internet-of-Crops® platform — compiles storage reports for various grain facilities around the world: temperatures, CO₂, moisture, fumigation status, all evaluated against a quality protocol which clients get to customize to their needs. We had three models generate the same report independently, from the same data, on the same day: GPT 5.3 as the baseline, Opus 4.8, and now Fable 5.

On the facts, the two Anthropic models agree almost perfectly — which silos are at risk, where fumigation is underway, which lots are running out of safe storage time. If all you wanted was an accurate scoreboard, either would do. The difference is in what each did once the scoreboard was filled in.

Opus 4.8 reads the data like a meticulous inspector walking the catwalk: silo by silo, rule by rule, every threshold checked, every alarm dutifully raised. It saw four silos running hotter than their neighbors, called them a cluster of probable biological hotspots, and ordered inspections. It saw outdoor humidity at 93% and gave the safe, categorical answer: do not ventilate. Sound, defensible, exactly what the protocol prescribes.

Fable 5 reads the same data like a storage manager with twenty harvests behind him. It noticed that eleven silos — different grains, different ages, different positions in the yard — were all pegged at the CO₂ sensor’s maximum at once, and asked the question an inspector wouldn’t: why everywhere, all at the same time? Its answer: the onset of summer warming the entire grain mass, a facility-wide seasonal effect rather than eleven separate emergencies. Then it went further and started second-guessing the sensors themselves. Silos under active phosphine fumigation are sealed, and sealed silos accumulate CO₂ — so the gas alarms there may be artifacts of the treatment, to be re-read after airing out.

One silo’s CO₂ had reverted from maximum to near-normal within twenty hours; Fable inferred it had probably just been ventilated and downgraded the alert accordingly. The hottest silos, it observed, were also the emptiest — and a low grain mass simply heats faster under a June sun. Even the humidity call was more operational: not “don’t ventilate,” but “ventilate in the low-humidity windows, watching the dew point.” It closed with a playbook — re-measure CO₂ 48 hours after ventilation, escalate if it stays high, unload lots in order of their remaining safe-storage clock. I admit I was impressed here.

In short: Opus tells you what the data says. Fable walks you to what the data means. The added nuances have been an unexpected delight.

The GPT 5.3 baseline is worth a paragraph, mostly as a measure of how far the floor has risen. It produced a fluent, competent report in a fifth of Fable’s runtime, and was the quickest to suspect that readings stuck at the sensor maximum might be a calibration issue. But it won’t — or can’t — go the extra mile. Combining multiple signals from different sensor types and deciphering them into actionable insights looks to be the realm of the latest frontier models.

Then we gave it the job nobody wanted

The second test came from Alex, who leads full-stack development in our group. It’s the kind of problem that software engineers are now happily delegating to AI. Configuration havoc.

Our simulation services sit on top of a handful of older software libraries, built in the era when Intel-style (x86) processors ran everything. Apple’s newer laptops use a different chip family — so our own developers’ MacBooks couldn’t run our own code. Anyone who has fought this class of problem knows why it lingers. It isn’t one decision, it’s dozens of interlocking ones: which old library do you upgrade, and to which version? Is the real culprit the library itself, or the way the services are packaged and wired to each other? And whatever you change has to keep working across four separate services at once. Each guess costs a slow build-and-test cycle to check, and after a few of those the creeping suspicion sets in that this will take ages and might not work at all. Every AI-assisted attempt up to and including Opus 4.8 had failed to produce something that runs.

Fable worked the problem for under an hour. At the end of it, the platform was running on an Apple Silicon MacBook for the first time in its existence. The actual fixes read like a checklist only hindsight makes obvious: two stubborn old libraries pinned to exactly the right versions, the plumbing between services rewired, and the startup scripts cleaned of commands that work on our Linux servers but quietly fail on a Mac. (For the engineers: TensorFlow for Rosetta compatibility, the worker image flagged linux/amd64, a Docker network bridge for inter-service networking, node-sass 4.11 → 4.14.1 for ARM build tools, and bash idioms swapped for macOS equivalents.) Five mundane decisions. The difficulty was never any single one of them; it was holding the whole picture in mind while iterating through slow, expensive feedback loops without losing the thread.

That is precisely the texture of most engineering busywork, and precisely what LLMs had not handled well until now.

The one I can’t tell you about yet

The third test I have to describe with the lights half-off. We have an initiative in internal demo that I believe will reshape how a large slice of U.S. agriculture sees its own market. At its core sits a hard data-acquisition problem: the information exists, but it’s scattered, inconsistent, and hostile to systematic collection. Exactly what you need agentic AI for — untangling a messy mosaic of live online data. Under Opus, coverage had plateaued; we’d assumed the remainder was simply unreachable without manual labor that doesn’t scale.

Switching the sweep to Fable didn’t just speed up the existing approach — it proposed approaches we hadn’t considered. Non-obvious methodologies, the kind a clever data scientist might land on after a week of staring at the problem, surfaced in the first sessions. Coverage is up a whopping 29% and still climbing. When the product surfaces publicly, I’ll write the full story, methods included. Maybe. For now, take it as the most commercially consequential of the four tests, and the one where the Opus-to-Fable delta was most obvious.

Revisiting my agentic workflow

Finally, a personal one. My maiden post on this Substack argued that agentic AI failed to impress me yet. I’ve kept running my personal productivity workflows ever since, partly as work and partly as a standing experiment: multi-step agents reaching across my email, our CRM, and our operations tools, assembling answers that no single system holds.

With Fable underneath, the flows are noticeably less brittle and noticeably more curious. Where Opus would retrieve the obvious artifact — the latest email thread, the top CRM record — and synthesize from it, Fable keeps pulling on threads: it cross-references a deal note against an ops ticket I hadn’t mentioned, notices that a contact’s silence coincides with a job change, retrieves context two or three hops away from where the question started. The nuance shows up in what it chooses to look for, not just in how it writes up what it found. I’m not ready to retract the maiden post — token consumption is still uncontrollable, and I still wouldn’t let an agent near anything irreversible without a human gate. But the gap to true autonomy is closing fast, and Fable shows a vector towards it.

Give Fable to the Masses, Please

Four data points don’t make a benchmark, and I’m suspicious of anyone declaring a new era off a weekend’s use. But the tests rhyme in a way I find more informative than any leaderboard. In the grain silo reports, the gain wasn’t accuracy — Opus was already accurate. It was the willingness to step back from the rules and ask what was really going on — including whether the sensors themselves could be trusted — then commit to a testable call with a plan to verify it. In the arm64 software port, the gain wasn’t code quality — it was patient troubleshooting across dozens of interlocking choices where every previous model lost the thread. In the project I can’t name yet, it was the model inventing new methods rather than executing ours. And in the agentic flows, it was retrieval driven by deeper business nuances. Four versions of the same underlying thing: judgment held steady over a long, messy context.

For those of us building physical AI — systems that sense, decide, and act on biological inventory in the real world — that’s the capability that matters. Nuanced insights and decision making which augments human perception rather than following preset rules.

So — is Fable really that good? Not the AGI promise yet, but a step toward it. The benefits for the business world have become tangible. For the good of humankind, and for those of us out to solve grand challenges — as we do for post-harvest supply chains and food security — let’s hope the security concerns get resolved and Fable is unleashed to the rest of the world again.

Is your visionary creation hindered by the current ban on Fable overseas? Do write me at sotiris at centaur dot ag. I’d be glad to compare notes.

Anyone who has watched harvest traffic stack up around a country elevator knows this is no mean feat. A faster receiving line can be the difference between a combine moving and a combine waiting.

But this is only half the problem.

A grain elevator can increasingly move grain without many people present. It still cannot reliably determine, in the fuller sense, whether a grain shipment should be accepted, discounted, blended, aerated, fumigated, rejected, held, sold, shipped, documented, or defended in a dispute. It doesn’t parse out provenance and traceability records. We have automated much of the motion around grain. We have barely begun to automate the judgment around grain.

A great example is CHS’s automated elevator in Herman, Minnesota. Feed & Grain described the site as approaching the long-imagined lights-out grain elevator, with RFID-enabled truck flow, automated probing, weighing, routing to three dump pits, scale ticketing, 1.4 million bushels (35.5K metric tonnes) of onsite storage, and 110-car train loadout. Farm Progress went further, calling Herman a totally automated and high-throughput elevator that never sleeps. The intended experience is almost serene: a trained farmer hauls in their harvested grain, arrives with the right identification, moves through kiosks and sensors, gets sampled, unloads, weighs out, and leaves. On a good night, the whole pass can run under ten minutes.

This takes some serious engineering, and the most revealing detail is which part was hard. It was not the scale ticket. It was the grain probe. Sampling is where the physical world starts misbehaving. Trailers differ. Loads settle unevenly. A probe has to hit representative points (kind of a needle in a haystack proposition), and the sample has to be good enough to support a grade, a routing decision, and a payment. Even in a highly automated lane, the instant the system touches the commodity itself rather than the paperwork around it, the problem gets messier.

Other sites point the same way. Cargill’s Illinois River terminal near Meredosia installed a comprehensive system in 2025 — RFID cards, automated weighing, probe-station workflow, moisture testing, grading, digital pit assignment, ticket printing. Kasa Controls’ rail terminal for MKC, a Kansas-based cooperative, emphasizes source-to-destination routing, commodity checks to reduce commingling (a serious issue when it occurs), automation logs, and train loadout cut from the traditional 16-20 hours to under 6.

Oklahoma State Extension lays out the architecture cleanly: identify the truck, weigh it, sample it, enter inspection results, calculate the grade, route the grain, and store the quality and destination data in the records. The same logic shows up on the commercial side, where CargillAg, ADM’s FarmView, Bunge’s grower portal, and ScoularView use apps to put bids, contracts, scale tickets, and settlements on a screen. The elevator is no longer just a scale house and a phone. It is becoming a connected workflow.

And yet the visible record from the grain majors and the automation vendors is far stronger on workflow than on judgment. The focus of digital elevators is on customer portals, automated receiving, and high-speed loadout, not on the biological and commercial state of grain received and stored. That gap is the whole story.

It matters more because the pressure to automate is not going away. It is actually becoming more severe. Some of it is labor. Feed & Grain has written plainly about the near-constant labor shortage in grain handling and feed manufacturing, driven by long hours and low awareness of the work. The demographic backdrop is no kinder: USDA’s Economic Research Service notes that nonmetro populations are aging, that the working-age population in nonmetro areas declined between 2010 and 2023, and that younger people keep moving toward cities. Its Rural America at a Glance series returns to the theme year after year.

So yes, automation means doing more with fewer people. But that is the least interesting version of the argument. The deeper issue is institutional memory. A good grain manager does not merely operate a facility. He or she knows which farmer’s corn tends to come in wet, which bins carry history, which loads deserve a second look, when a discount is technically correct but commercially explosive, which buyer will tolerate which quality, when a blend is clever and when it is reckless, and when the weather has turned a storage decision into a race. When that knowledge retires faster than it can be replaced, the facility does not just lose hands. It loses judgment.

Harvest makes all of this visible. The bottleneck is not only a line of trucks. It is a line of unresolved decisions. A farmer may have several delivery points, different bids, different hours, different dump speeds, different discount schedules, different drying capacity, and a weather window indifferent to anyone’s preferred timetable. ALCIVIA’s Hager City facility, with three 15,000-bushel pits, was built explicitly around speed for farmers used to long waits. GRAINLAND Cooperative in Illinois cut peak harvest waits from roughly 30 minutes to about five after upgrading receiving and drying.

Put these threads together and the shape of the achievement comes into focus. The industry has built elevators that can take grain in faster, later, and with fewer people than at any point in their long history. It has automated the choreography almost completely. What it has not automated — what it has, if anything, begun to quietly dismantle by retiring the experienced people who used to supply it — is the judgment: what the grain actually is, what it is worth, what should happen to it next, and who gets to decide. That is the harder and more consequential half of the problem, and it is the half that decides who captures the value at the pit.

It is also where this essay turns from admiration to vision. In Part 2, I will follow the judgment problem to the precise moment it becomes money and power: the discount, and the mechanics of the basis trade; and what happens when the person in the scale house is replaced by a machine. Then, finally, what a genuinely intelligent post-harvest system — physical AI that understands the commodity and not merely the choreography — would actually have to do to earn the trust we are about to hand it.

Stay tuned.

Interactions at an event like this are partly AI-sourced. I can now reveal that the agentic flow mentioned in my previous Substack helped me home in on the innovators I could not miss meeting. But they are also wonderfully serendipitous. Like when I bumped into Konstantinos Aretakis, a bright young fellow working with Tools for Humanity in San Francisco – a hive of forward-minded creators developing the Orb, an apparatus promising to solve an increasingly tough problem: how to set humans apart from synthetic intelligence.

In our highly inspiring exchange, and ignoring a 30+ year age gap with my interlocutor, I brought up how this reminded me of Blade Runner. I was hoping Konstantinos might have seen at least the remake starring Ryan Gosling, if not the original where Harrison Ford was tasked with essentially what the Orb is now aiming to do – figuring out if Rachael was human or synthetic.

I was also impressed with how Dexory, represented by its co-founder and Chief Commercial and Product Officer Oana Jinga, deploys robots and AI to map labyrinthine warehouses around the world. My team at Centaur is out to analogous feats in the messy world of grain warehouses, although our sensors and AI are trained on insects, CO₂ gradients, and the occasional reminder that agriculture does not read product manuals. Dexory’s traction is phenomenal, no doubt thanks to excellent execution, but also due to underserved needs in the physical world that AI can cover when it starts understanding space, inventory, and motion.

Serendipitously, I had the opportunity to attend the fireside conversation between Christian Bach (Netlify), with Demetri Kofinas of Hidden Forces, on Europe vs. Silicon Valley: Two Models of Innovation. What an engaging thought leader Christian is. I almost shouted “yes!” at parts of that conversation, which would have been undignified, even by ancient Athens standards. Having operated on both sides of this divide, I very much wish the virtues of each could be transplanted to the other: Europe’s depth, patience, and scientific seriousness; Silicon Valley’s velocity, ambition, and extraordinary tolerance for things that sound unreasonable until they become inevitable. But is the world better off with those virtues remaining distinct and at arm’s length?

I was also elated to meet George Sarakostianos and Manos Xanthakis of Agritrust whom I “charmed” to move on and win 2nd place distinction at the pitch event! From the plains of Thivai, this duo of political economists has set out to make sane one of the toughest challenges in agriculture: predictability and continuity of supply in contract farming. Gentlemen, you approached me for advice, and I hope I did a decent job. But here is another piece of it: do not do many more pitch competitions. Keep building product for the benefit of farmers in Greece and elsewhere.

What stayed with me from Panathēnea was not a single panel or meeting, but the convergence of makers around physical AI. At Centaur, we are building post-harvest intelligence for grain and commodity supply chains. We are stoked to see AI leaving the chat window and entering the world of objects, facilities, identity, logistics, and risk.

Interested in a Part 2 of Panathēnea with movie references? Like this and coax me to continue. Also drop me a note at sotiris at centaur dot ag. Synthetic can also be beautiful – ask Rachael above.

This is not a criticism of agentic AI technology in its infancy. It is a criticism of the paradigm under which we are deploying these tools. We are asking AI agents to automate workflows inside digital systems designed for human eyes, human clicks, and human patience. It works, sometimes brilliantly, but it also feels like using the Hubble telescope to read a road sign. Should conference websites ship with MCP connectors already? I am increasingly convinced the answer is yes.

Personal perspective

When we first started designing the Internet-of-Crops® platform almost ten years ago, “agentic AI” was not a term of art. Nor had the world become familiar with World Models as the next frontier in humanity’s pursuit of synthetic general intelligence. Our starting point was more practical. A large share of food waste happens after harvest, in the silent interval between field and processing: inside silos, warehouses, fumigation chambers, vessels, containers, and supply chains still operated with little premeditation and too much habit. Our vision was to make that invisible post-harvest layer measurable, predictable, and ultimately controllable, so that stored commodities could move through the world with less waste, less energy, less chemical overuse, and better economic outcomes.

If we could combine sensor data with simulation, we could do more than simply monitor grain. We could understand what was happening inside a silo: where moisture was moving, where heat was building up, whether insects or mould were becoming active (looking at CO₂ buildup), and whether a fumigation was actually working. From there, we could predict what was likely to happen next and test the best response before acting in the real world. For example, the system could decide when running the aeration fans would protect grain quality while using less energy. With the Internet-of-Crops®, that decision can be recommended to an operator with the tap of a button, or even carried out automatically by the platform.

Video: A fluid dynamics simulation of a grain silo

Missing the point
So this is where much of today’s discussion about workflow automation misses the point:

• In factories, an agent should not watch an HMI screen to infer that a line is drifting out of tolerance. It should query the production model, maintenance history, sensor state, and quality constraints, then simulate whether slowing a conveyor, changing a setpoint, or scheduling downtime protects throughput.

• In supply chains, an agent should not just scrape shipment portals. It should reason over inventory, port congestion, temperature excursions, contract terms, and risk tolerances before proposing a reroute or release decision.

• In medical workflows, the goal should not be an AI that clicks through records faster, but one that can assemble patient context, clinical guidelines, lab test files, and care-team permissions into a safe recommendation pathway.

What this approach teaches us is straightforward:

1. Automation is not the same as teaching software to imitate a user sitting at a keyboard (this is also token-intensive and gets expensive pretty quickly).

2. Real automation requires a model of the domain, trusted interfaces into the systems of action, and a digital model that tells the agent whether the world behaved as expected.

3. The next generation of platforms will expose their internal state cleanly, describe their capabilities in machine-actionable ways, and provide safe rails for autonomous decisions.

4. MCP connectors and similar protocols are not just developer conveniences. They may become the plumbing through which AI stops pretending to be a human user and starts becoming a genuine operational layer.

At Centaur, we came to this conclusion through an objective to eliminate wasteful habits in grain handling. Others will arrive there by reimagining their own path to more efficient workflows. The trajectory may differ, but the conclusion is the same: agentic AI does not need more elaborate ways to look through the window. It needs plumbing built for it.

If you are thinking about agentic AI beyond screens and unleashed inside your plant, or wonder about the future of post-harvest supply chains, write me at sotiris at centaur dot ag. I’d be glad to compare notes.