1783
A scruffy, post-war port
If you’d stood on a muddy New York street in 1783 and confidently predicted that this scruffy, post‑war port would become the greatest city in the world, people would have laughed in your face.
They would have been right to.
New York, at that point, was a mess. Loyalist baggage. A small population. Overshadowed by places like Philadelphia and Boston. If you were placing long‑term bets based on the data in front of you, New York would not have made your shortlist.
And yet, that scruffy port becomes the epicentre of global trade. Not because someone declared it so, but because a set of extremely unglamorous, highly technical decisions compounded over time.
This is a story about one of those decisions, the person behind it, and why his ghost is walking around the AI and ecommerce world right now.
It’s a story about Canvass White, leaky canals, and why everyone building “agentic commerce” needs to be a little bit more like a 19th‑century engineer and a little bit less like a 2020s keynote deck.
The Man
Introducing Canvass White
Fig. Of course, this is a less than 2MP lens camera from the age, so it a bit pixelated.
1800s
More wall than welcome mat
Roll the clock forward from 1783 to the early 1800s. Extract following from Bill Bryson.
New York has exactly one structural advantage: geography. It has a plausible route through the Appalachian Mountains to the west, where the farmland and futures are. The problem is that “plausible” and “practical” are not the same thing. The Appalachians are more wall than welcome mat. Moving goods overland is ruinously expensive.
At the time, farmers west of the mountains found it cheaper to ship down the Ohio and Mississippi to New Orleans, then all the way around Florida and up the Atlantic coast, than to drag their stuff a few hundred miles over dirt and rock. Three thousand miles by water beat three hundred miles by wagon.
1817
They nicknamed it “Clinton’s Folly”
Enter De Witt Clinton, mayor of New York and future governor. Clinton proposes a solution that is so audacious it earns its own punchline: a canal running across the entire state to Lake Erie, connecting New York to the Great Lakes and the rich hinterland beyond. People nickname it “Clinton’s Folly.”
The reasons are obvious:
- No one has built a canal this long or this technically demanding, anywhere.
- The route is 363 miles of wilderness. The canal must be cut with picks and shovels, forty feet wide, with eighty‑plus locks to manage elevation changes.
- The gradient on some stretches has to be tuned to about an inch per mile. That is a “measure twice, dig once” kind of precision.
- America does not have a single native‑born engineer who has actually built a canal. Not one.
Clinton’s big idea is essentially: “Let’s attempt the most ambitious canal project in the world with no experienced canal engineers, across difficult terrain, with huge political opposition and no federal support.”
If you’re thinking, “This sounds suspiciously like every AI platform roadmap I’ve seen in the last 18 months,” you’re already seeing where this is going.
The state of New York decides to do it anyway.
Maybe they were right. Results since are decent.
The Team
Three judges and a schoolteacher
They appoint four men to oversee the project. Three are judges; one is a schoolteacher. Their shared expertise? Some surveying experience and a willingness to read a lot, ask questions, and experiment like hell. Together they somehow design and run what becomes the largest engineering project the young United States has ever seen.
But early on, a very specific, very boring problem almost kills the whole thing.
The Crisis
The canal leaks
Not metaphorically—physically. Dirt and stone will not hold water on their own. To keep the canal watertight, they need hydraulic cement in staggering quantities: something like half a million bushels. If the water seeps away at any point, large stretches of the canal become useless.
No watertight cement, no canal. No canal, no New York.
This is where Canvass White walks in.
The Fix
The guy who fixed the leak
Canvass White is a young canal employee who volunteers to go to England at his own expense to figure out how grown‑up countries build canals that don’t drain themselves. He literally walks around Britain studying towpaths and locks, logging about two thousand miles on foot, taking notes on everything—design, materials, maintenance.
He discovers that a material called Parker’s Roman cement, which had been a disaster for one grand architectural project (Fonthill Abbey), is surprisingly good as a hydraulic cement. It is not great for buildings, but it is excellent for waterproof mortar.
White returns to America with a working understanding of the science behind adhesion and the practical realities of canal engineering. He doesn’t copy and paste the British approach. He experiments with local ingredients and develops a compound that works even better than Parker’s cement in this specific context.
The result is a hydraulic cement that:
- Can be produced from American materials.
- Works as a reliable, cheap, scalable way to keep the canal from leaking.
- Enables the project to complete early and under conditions that would otherwise have been impossible.
It is, quietly, one of the most important technological breakthroughs in early US history. It literally holds together the infrastructure that will make New York what it becomes.
The Injustice
Four cents a bushel
White patents his cement and is supposed to receive a royalty of four cents per bushel. In practice, manufacturers ignore him, grow rich, and drag him through the courts. He never sees the upside his invention deserves. He eventually dies poor and largely forgotten, while New York explodes into a commercial powerhouse.
1825
The canal opens
People describe the running lights at night as looking like fireflies on the water.
The canal opens in 1825. Boat traffic is so heavy that people describe the running lights at night as looking like fireflies on the water. Shipping costs fall off a cliff: moving a ton of flour from Buffalo to New York drops from around 120 to 6, and transit time drops from weeks to days. New York’s share of US exports and its population both skyrocket over the following decades.
If you trace the line from that cement to the financialisation and urbanisation that follow, you can make a serious argument that:
- Canvass White’s cement doesn’t just enrich New York.
- It shifts the balance of economic power in North America.
- It contributes, indirectly but materially, to the United States becoming the dominant economic player of the 19th and 20th centuries.
And yet, almost no one talks about Canvass White.
Infrastructure has a PR problem. A bit like AI IMO.
1994–
Ecommerce’s 30-year canal loop
Now, pivot to ecommerce.
We’ve spent roughly three decades building what, in canal terms, are a lot of boats and a lot of docks, with some very inventive paint jobs.
The basic loop has been weirdly stable:
- A human decides they want something.
- They go to a site or an app.
- They search or browse.
- They pick a merchant.
- They pay.
- They wait.
Innovation has mostly cycled around:
- Where we capture attention: search, banner ads, email, social feeds, influencers, short‑form video.
- How we show the goods: static catalogues, tabbed interfaces, “miles of aisles,” personalized recommendations, mobile apps, live shopping, endless DTC variations.
- What rails we run on: cards, then more cards, then digital wallets and BNPL, layered onto legacy banking and logistics.
If you scroll the screenshots of early Amazon, NetMarket, Pizza Hut’s PizzaNet, eBay’s AuctionWeb, Snowdevil (the Shopify seed), Magento admin, BigCommerce dashboards, through to today’s SaaS storefronts, the UI changes, the copy changes, the buzzwords change. But the stack underneath is essentially the same bargain: “We’ll put all the work in front of you; you can do it faster and maybe slightly cheaper than in 1998.”
We’ve had waves that felt like revolutions:
- Search and SEO.
- Social commerce and Facebook pages.
- Marketplaces swallowing independent stores.
- Mobile apps and app‑based loyalty.
- The DTC boom and bust.
- The creator economy trying to become a commerce rail.
They were, in retrospect, arbitrage windows:
- Arbitrage on new attention surfaces.
- Arbitrage on new forms of data.
- Arbitrage on new capital flows and subsidised logistics.
But in all these loops, the fundamental thing remained true: a human is the orchestrator. We built better signs and smoother floors; we did not change who was pushing the trolley.
Now
Agentic commerce: new canal or new cement?
Now we are told that “agentic commerce” is here.
On the surface, it is very easy to feel like this is just another fad. Another three‑letter acronym, another parade of demos where someone chats to a bot that magically books flights, assembles outfits, reorders groceries, reconciles expenses, and probably walks your dog.
If you have been around long enough, a part of you wants to say: “Sure. We’ve seen this movie. Call me when it ships.” Authors note: Resisted to write “WIZMO” here.
But if you look underneath the demos, something genuinely different is trying to happen.
The agentic vision, at its most interesting, is not “chatty frontend” but “different stack.” In a mature version of this world:
- Users express goals and constraints (“I need to kit out my home office for under €1,000, sustainably, before month‑end, and I hate assembling furniture”).
- Agents translate that into structured tasks: searching, negotiating, splitting orders, handling returns, respecting your preferences and obligations.
- Protocols exist that let agents talk to merchants, payment systems, logistics providers, and each other with shared expectations around identity, authorization, liability, and data sharing.
- Instead of pushing a trolley down digital aisles, you delegate the entire errand to a trusted system that routes through multiple merchants, regions, and constraints on your behalf.
This is not a new canal.
The protocols may be the most needed part for us Average Joes. But the rise in Agentic is fuelled through adoption and money. ROI not yet needed. Yet.
The rails—payments, logistics, SKUs, even many platforms—are still there. This is a new form of hydraulic cement:
- It sits between layers that already exist.
- It solves the “leakage” between them: the endless friction where intent, catalog, payment, and fulfilment fail to line up.
- It’s invisible when it works but transformative when it scales.
The thing to pay attention to is not “Does this chatbot feel smarter than last year’s chatbot?” but “What are the protocols and infrastructure that make it safe, reliable, and economically meaningful for agents to transact on our behalf?”
Because that’s where the arbitrage is.
Windows
The new arbitrage windows
The Erie Canal created a massive, time‑bounded arbitrage: for a while, New York had an extraordinary cost and speed advantage for moving goods from the interior to the world. Eventually others caught up—railroads, competing routes, new ports—but for a critical window the differential was world‑shaping.
Agentic commerce, if it works, will create similarly time‑bounded arbitrage pockets.
A few that are starting to become visible. They are not guarantees of the exact path forward but maybe that is the path forward. Non exacting:
Attention arbitrage 2.0: from feed to agent
Today, ecommerce arbitrage is overwhelmingly about:
- Winning the feed (social, search, marketplace).
- Extracting more conversion per impression.
- Squeezing shipping or acquisition costs a bit further.
In an agentic world, more and more discovery and evaluation shifts from human attention markets into agent‑to‑agent negotiation.
Your buyer’s agent might:
- Maintain a live context of the user’s preferences, constraints, and obligations.
- Query multiple sellers’ agents or APIs in the background.
- Filter, rank, and negotiate based on rules the human never sees directly.
The arbitrage window is: being among the first environments that agents prefer to deal with.
That means:
- Clean, well‑documented product data.
- Programmatically accessible policies (returns, warranties, sustainability).
- Stable, agent‑friendly protocols for browsing, ordering, and resolving issues.
It is SEO before marketers realise SEO is a thing: a world where the few who expose “agent‑ready” surfaces enjoy outsized inbound demand because everyone else is still optimizing Instagram Reels.
Complexity arbitrage: friction as a moat
Human shoppers are allergic to complexity.
Ask them to juggle:
- Tariff rules.
- Multi‑carrier fulfilment.
- Complex benefit schemes.
- Cross‑border tax scenarios.
They will either give up or default to Amazon.
Agents, on the other hand, can eat complexity for breakfast, as long as:
- The rules are codified in machine‑readable form.
- The incentives are aligned.
- The protocols allow negotiation and exception handling.
The arbitrage here is to be the first to turn complexity from a tax into a moat. If your system can:
- Assemble optimal cross‑border bundles.
- Route around stockouts politely.
- Navigate benefit entitlements.
- Handle returns intelligently.
…all without interrupting the human, you are suddenly playing a different game.
Regulatory and protocol arbitrage: writing the locks
Someone is going to make the early decisions about:
- How agents authenticate.
- What they can consent to.
- How disputes are logged and resolved.
- What data they can carry between contexts.
These are not just technical choices; they are policy. They define:
- Who bears risk.
- Who sees what.
- Who gets to say “no.”
The companies, consortia, and regulators who step up here are effectively deciding where the “locks” go in the new canal network, who controls them, and how much friction they create.
For a while, there will be gaps and mismatches between jurisdictions. That creates arbitrage for players competent enough to operate safely in the grey zones while the standards catch up.
Eventually, as with seals on containers and SWIFT in banking, these things standardise. The arbitrage becomes less about being the only one and more about being very, very efficient on a shared set of rails.
But right now? The cement recipes are still being written.We need it, we need control because we used to be able to rely on PEST. Now we are being shaped by a pest
But ecommerce has got to change, find a new path forward.
Our greatest ecommerce achievement is not really a shouter-outer is it?
Method
Walking the canal
What Canvass White did well was not reading about canals.
He walked them.
He:
- Saw where they cracked and where they held.
- Talked to the people maintaining them.
- Observed how theory met mud.
- Then came home and iterated, with local constraints and materials.
There’s a version of this mindset that applies painfully well to AI in commerce.
If you are serious about agentic commerce, the job now is not “buy the hottest tool and wait for magic.” It is to walk your own canal:
- Follow an order from intent through search, recommendation, inventory, fraud checks, shipment, delivery, returns, customer service.
- Instrument where the water “leaks”: confusion, abandonment, disputes, write‑offs.
- Look at which of those leaks are caused by missing or messy infrastructure rather than missing “content.”
Then, like White, you need to experiment with “cement”:
- Product schemas your agents can rely on.
- Policies written in a way both humans and machines can understand and enforce.
- Authentication and consent flows that can be delegated without surrendering control.
- Escalation rules and kill switches when things go wrong.
You will absolutely discover problems halfway through the build that you could not have anticipated up front.
White did not solve hydraulic cement in a lab and then, once it was perfect, give the green light for the canal. The canal forced the cement into existence.
In the same way, the only way to build meaningful agentic commerce is to run real, bounded workloads and let the failures show you where the missing infrastructure lives.
If you try to design it all in PowerPoint, you are back to judges designing canals without ever seeing one.
Theses
The Cluetrain Manifesto… with agents
When The Cluetrain Manifesto came out in 1999, its big thesis was simple and electrifying: “markets are conversations.” It argued that the web was reconnecting buyers and sellers as real humans, after decades of broadcast, one‑way, mass‑media marketing.
That line has aged surprisingly well, but the environment around it has not.
Today:
- Conversation is happening in more places than ever—feeds, comments, DMs, reviews.
- And yet, weirdly, the actual work of commerce often feels less human, not more.
- There is more spam, more sludge, more performative engagement, more algorithmic optimisation, less genuine listening.
So what happens when we throw agents into that mix?
One answer is dystopian: markets become machine‑to‑machine combat, with no genuine human in the loop. Endless optimisation for margin and engagement, opaque incentives, and real people reduced to behavioural exhaust.
Another answer is more hopeful: markets become conversations between systems that are explicitly and transparently representing human interests, under human‑set guardrails.
To get closer to the second scenario, we probably need some updated theses—call it Cluetrain 2.0 for agentic commerce.
Something like:
- Markets are conversations between systems on behalf of humans. It is no longer just brand to person, or person to person. It is increasingly system to system—agents negotiating, querying, and transacting—but those systems are not neutral. Someone designs them, funds them, and benefits from them. That needs to be explicit.
- Every agent carries someone’s intent—and someone’s risk. When your “shopping assistant” makes a decision, it is acting under some set of incentives. Maybe that’s you; maybe it’s the retailer; maybe it’s a hidden affiliate partner. That chain of intent and liability must be visible, auditable, and challengeable.
- Protocols are policy encoded. The way you structure access, ranking, identity, and payments in your APIs and protocols is a value judgment. It decides who gets a shot, who gets sidelined, and who gets exploited. If you are not in the room when those decisions are made, they are still being made—just without you.
- Arbitrage without alignment is extraction. Using agents to exploit gaps in user understanding, regulation, or system design will work… until it doesn’t. It creates short‑term gain and long‑term fragility: backlashes, bans, broken trust. Sustainable arbitrage uses new capabilities to create real surplus for all parties, not just invisible rent for the middleman.
- Guardrails must be legible to people and machines. If your safety, fairness, and governance principles exist only as legal text, agents can’t follow them. If they exist only as machine rules, humans can’t scrutinise them. They must exist in both forms, and they must be reconciled.
- Invisible infrastructure demands visible accountability. As more of commerce disappears behind the phrase “my assistant handled it,” the stakes of getting the infrastructure right go up. We will need more transparent logs, stronger audit rights, and independent oversight, not less.
- The right guides are canal‑builders, not carnival barkers. If you have to choose between putting someone on the steering committee who knows how to run a critical payments system and someone who knows how to juice DAU with dark patterns, choose the former. The next decade of composite AI and commerce will look less like a marketing campaign and more like an infrastructure project. Build your leadership bench accordingly.
- If we are not building resilience, we are repeating PTSD. The last decade of ecommerce has left a lot of people—consumers, merchants, even teams—feeling tired, guilty, and trapped in systems optimised for “more” rather than “better.” The point of AI here should be less cognitive load, healthier margins, and more robust systems, not another turn of the anxiety screw.
This is the piece most missing from the “wow” demos: a set of shared principles that outlive the hype cycle.
Next
So, what now?
If you are reading this as:
- A brand trying to work out if it should “do something with AI.”
- A SaaS provider staring at your roadmap.
- A retailer trying to keep your head above water.
…there’s a simple, unglamorous answer.
Be Canvass White.
- Walk your canal.
- Learn the boring, structural details of your stack.
- Find the leaks.
- Invent the equivalents of hydraulic cement: protocols, schemas, guardrails.
- Accept that you may not get a hero’s reception for doing it.
Because in hindsight, the people who matter most in these transitions are not the ones who name the era. They are the ones who quietly make the era possible.
We are not starting from zero. Nothing about this is entirely new. AI is not magic; it is another layer in a very long history of humans building tools to move things—goods, money, attention—from one place to another.
The opportunity now is to treat this not as another campaign but as infrastructure. To ask, with a straight face: “What if this is our Erie Canal moment? What if the real work is to fix the leaks?”
And then, shoes on, notebook in hand, start walking.