The third in a series of Silicon & Steel conversations with the founders building what's next.
Welcome back to Silicon & Steel conversations. The last two episodes stayed close to the physical world, first the empty trucks on the highway, then the blind spots inside the box. This time we step up a layer, to the AI agents now being pointed at supply chains, and the harder question underneath the hype: can you actually trust them to run anything?
My guests are Ben Roome, co-founder and CEO of Eloso Intelligence, and Drew Winget, co-founder and CTO. Roome spent twelve years in responsible AI and governance before supply chain; Winget spent more than a decade in software before startups; their third co-founder, Jake Metcalf, could not join but shaped much of the thinking here. Eloso builds AI agent harnesses: the structure wrapped around an AI agent that makes it work reliably in the real world.
The full conversation is on YouTube, linked below. What follows is the short version.
What's Inside
Let’s get into it.
1. The Problem: The Pilot That Breaks
Every company is being told to deploy AI agents or fall behind. Ben Roome's pitch deck opens with the opposite claim: "Problem. Your AI agents don't actually work." He wrote it on purpose, because it is what he hears from nearly everyone. Agents work beautifully in the demo, and for the first week or two of the pilot. Then their processes degrade and they start making mistakes no one is watching for. The failure is not intelligence. It is consistency, and in supply chain, inconsistency is expensive.
“We hear horror stories of agents sending a bunch of POs out to suppliers that shouldn't have gone out. Now these all have to be rescinded; they've undermined the trust between the company and the supplier. If you don't have a harness, you cannot be certain your agent will not start making serious errors.”
2. Five Things That Stuck With Me
i) The harness is the wisdom, not the intelligence
The model is raw capability. The harness is the human judgment wrapped around it, the rules, context, checks, and fallbacks that make it safe to run unsupervised.
“Intelligence is the ability to solve problems, and wisdom is the ability to avoid problems. An LLM is artificial intelligence. It is not artificial wisdom.”
ii) An agent is a trigger plus a memory
What separates an agent from ChatGPT, or from old automation? The trigger and the memory. ChatGPT waits for you to type; an agent fires on a clock or an event, and carries context across all of its past work.
“What makes something an agent is this hybrid of automated triggers, memory, and an LLM. Old automations are brittle. Now that LLMs are here, it is possible to do this much more flexibly.”
Winget's plainest framing of the harness: "You are essentially writing a process manual and giving tools to a robot employee."
iii) The agents don't work until the harness does
Roome splits the harness into operational components (rules, tools, memory, context) and control components (verification, permissions, fallback paths, audit logs). Skip the controls and you get the horror stories.
“In a lot of cases, your agents don't work until you have really put energy into building the harness properly.”
iv) The future is a swarm, and swarms breed chaos
No company ends up with one agent. It ends up with a workforce of them, from different vendors, inside and across companies. Orchestrating that swarm becomes the real work.
“Companies are going to have hundreds to thousands of agents. If you had the ability to hire a thousand people for ten dollars, I'm pretty sure you would do it.”
"The more agents you introduce to a system," Roome adds, "the more chaos you introduce as well."
v) The future is here, just not evenly distributed
Asked where agents fit alongside containers, ERP, and e-commerce, the founders reframed it: the story is not one invention, it is the widening gap between the companies that have adopted and the ones that have not.
“I've written more code in the last six months than in the last thirteen or fifteen years of being an engineer. And I've read almost none of it. That is just going to happen to every industry.”
3. The Broader Implications
Notice who is making this argument. All three founders share a background in philosophy, and Roome and Metcalf came from AI ethics and governance. Eloso is not a faster-model company. It is selling trust and consistency, the unglamorous infrastructure of making a powerful thing behave, on the assumption that intelligence keeps getting cheaper while the wisdom around it stays scarce. Roome's image for the industry says the rest: every company's supply chain is a Rube Goldberg machine, and the whole global system is a giant one built on top of little ones. Drop unsupervised agents into that, across company lines, and a governance layer stops being abstract.
4. The Takeaway
The bet, in one line: as intelligence commoditizes, value moves to the wisdom around it, and whoever owns the harness owns the reliability. The warning folded inside it is the useful part for anyone in supply chain today. The adopter divide is real and widening; the companies learning to run agents well are compounding a lead the others cannot yet see. Catch up too late, in Winget's phrase, and it will feel like you are running into time travelers.
5. The Last Word: Why to build in Supply Chain
I end these conversations the same way: what would you tell someone young enough to spend a whole career on this? Roome, whose eyes would have glazed over at "supply chain logistics" four years ago, has since fallen for it, everything you own reached you because the chain was working, and everyone remembers what it felt like when it stopped. His advice to students is to learn what agents can do now, and to founders, the old rule: find a real, mission-critical problem people have budget to solve. Winget's closing thought may be the one to carry furthest.
“For the first time in our lives, the ability to articulate your desired outcome is real leverage. Just being able to say what you want means you can get ninety percent of the way there using AI.”
Don't get attached to any one tool, he says, because everything is changing too fast to hold on to. The key skill, in his words, "is going to be your ability to adapt."
Which is as good a reason as any to keep watching this space, and this series.
Subscribe if Silicon & Steel is your kind of thing.
More conversations are on the way, and my thanks to Ben and Drew for being on this talk show.
Advertisement
Domain Names + Web and Email Hosting You Need
Still paying GoDaddy or Namecheap prices? Porkbun sells most domains at cost for low, transparent registration and renewal pricing with no nonsense. Get free features like WHOIS privacy and SSL certificates, plus real human support 24/7, 365 days a year. Save $1 on your next domain name now.
Hiring in a new country? Read this first.
Every country has different employment laws, payroll requirements, notice periods, and benefits expectations.
Oyster's Global Hiring Guides help you navigate the details, avoid surprises, and hire with confidence—wherever you're growing next.
Ads Ends
