Multi-Agent Systems in an AGI Future

Watch on YouTube • Listen on Spotify • Listen on Apple

There’s a tidy old question economists like to ask at dinner parties when they’ve run out of wine: Why do firms exist? After all, if markets are efficient, shouldn’t it always be cheaper to outsource your operations to external providers? Ronald Coase’s Theory of the Firm proposed a brutally practical answer—because the cost of using the market (finding vendors, negotiating, coordinating) is often higher than doing it inside the org. Lower those outside costs and the boundary of the firm shifts outward like a tide.

This idea matters because while we’ve clearly established that AI can produce individual pieces of code or content materially cheaper than human beings, we have yet to show that the coordination costs actually decrease within a firm. If you believe in the theory that AI allows companies to be much smaller than before, you are actually saying that you think internal coordination costs are going to dramatically decrease. Otherwise every additional meeting gets exponentially more expensive as your staff get more and more leverage out of their time spent.

Allow me to ask this question in another way: What happens when “the market competitor” isn’t another vendor or headcount but a meter—AI agents that can log in, click, remember, and obey rules? If metered compute + a little supervision costs less than new payroll or new vendors for the same reliability, the rational move is unfancy: don’t hire someone—meter an AI agent.

This move is the through‑line of my conversation with Flo Crivello (Lindy). His company creates horizontal agents that automate all the annoying, ticky tack work that makes companies move slow. For the first few years, Lindy’s product was pretty good but not amazing. But over the last 6 months, that has totally changed. The models are finally good enough that the cost of coordinating agents is less than actually doing the work yourself. What changed? Three things:

* Computer Use finally started working.

* Horizontal agents became generalizable enough to flex across the common tasks that all companies do.

* “Lindy runs on Lindy,” Flo estimates that maybe in about two years he’ll spend more on tokens then he will on payroll. (Wild). While the tech’s improvement is key, the culture at Lindy and other companies moving towards being AI first allowed for this all to work.

I’ve put my personal takeaways below including how I’m thinking about changing my own business, the quotes from Flo that stuck out to me, and how to think about vertical versus horizontal AI solutions.

1) Computer Use: when software grows hands

Lindy gives each agent a persistent cloud computer—a real screen, cookie/session memory, and a takeover button when the flow gets weird. If a person can do it in a browser, the software can do it now.

“As part of Lindy 3.0, what we released is computer use, which massively advances us along the capability axis. So basically, we are giving each agent its own computer in the cloud. So it can do anything that you can do on a computer. And Agent Builder… it’s literally we’ve used Lindy to build this agent.”

The goal with computer use and agents isn’t to make them smarter, it’s to make them more reliable.

“If you put two doors side by side, and one of them is an automatic door, and the other one is a manual door, and the automatic door works only 98% of the time, people are going to use the manual door 100% of the time. … It’s got to work extremely reliably in order for you to use it as your default.”

Computer Use also bulldozes the “we don’t have an integration for that” backlog by changing the surface area of what’s possible:

“We’ve got almost 7,000 integrations at this point on Lindy… and yet we realized it’s never going to be enough. We’re always bottlenecked… computer use… once and for all gives us access to all the [integrations] we need.”

And yes, it’s already past the toy phase:

“I used it an hour ago… I ordered vitamins and creatine because I’m running out. And I just asked ‘buy them on Amazon for me’. And she did.”

“We fixed [the re‑enter password problem]… it persists your logins—just like your Chrome browser… For me, buying stuff on [Lindy] is literally just sending it a message… and it just buys it.”

Under the hood, the “hands” are paired with scaffolding—verifiers, retries, and policy:

“We are using [Claude] Sonnet, which is cracked at computer use… we have built a lot of scaffolding around computer use to make it better… we’ve got evals that… we are a lot better than the state of the art.”

I have found that many software companies are underestimating how important computer use is. I think in about 12 months, many applications will be due for a reckoning on how different their users are going to be.

2) Horizontal vs. Vertical: a procurement rule you can say out loud

The honest version:

* Horizontal platforms feel like a 6/10 at everything.

* Vertical tools are a 9/10 at one thing.

So why pick the 6/10? Because seams have a tax—every extra point solution imports reviews, contracts, training, dashboards, renewals, and one more place for the workflow to snap. Coase would call these transaction costs; your CFO calls them “why does this take nine people.”

Flo’s frame is pragmatic and true to how ops actually work:

“AI agents are a new category, but the category that they fit most in is iPaaS… the winners have been extremely horizontal. UiPath, Workato, Zapier…all of these guys have always been extremely horizontal.”

“Being horizontal makes you sort of 6 out of 10 at everything… In the verticals… this player is going to be 9 out of 10… So why buy the 6/10 when there’s a 9/10 next door?

Reason #1: the use case isn’t important enough… you don’t want 5,000 accounts.

Reason #2: the vertical tool may not do exactly what you want… The moment you color outside the box, it’s not going to support your workflow… 99% of the time, a vertical player does not support that exact workflow… 1% of the time… they actually want the cookie cutter… then we tell them to buy the vertical.”

Who buys what, in reality:

“We’re really more targeting SMBs… sweet spot is 20 to 200 people… Most of the time it comes from the top. The board: ‘what’s our AI strategy?’ The sheet rolls downhill… we jump on a call with Sales, Support, Ops and they tell us the workflows they want to automate.”

And how demand actually shows up:

“It’s been mostly inbound… We also come up in ChatGPT a bunch… double‑digit percent of our traffic comes from ChatGPT.”

The rule is essentially buy vertical when the job is standardized and deep; buy horizontal when the real workflow colors outside any single product’s neat box and changes weekly. That Lindy is mostly SMB made sense to me too, the smaller they are, the easier it is to re-architecture your business to be AI first.

3) “Lindy runs on Lindy”: tokens vs payroll

Flo runs a ~45‑person company with thousands of his own agents. He tracks something you can steal because it’s beautifully unglamorous: Token Spend vs Payroll.

“We punch above our weight in revenue per employee… and we’re using Lindy a whole lot. It’s absurd how much we use Lindy. We’ve got thousands, like literally thousands of Lindy’s just for the company… half of the company runs on AI agents.

We are actually tracking our token spend compared to our payroll spend… the lines will cross at some point…My hunch is in two years or so I think they’ll cross.”

What does that mean for who you hire? Fewer biz ops and back office staff; more Agent Ops (rules, observability, SLOs) and Workflow Engineers (turn messy runbooks into reliable click‑paths). The coordination work doesn’t vanish; it moves—from email and muscle memory into rules you and an AI agent can read.

Staffing posture matters, too:

“I love young people… It’s powerful to pair really senior people with really, really younger people—young bring energy and innovation; seniors anchor them to reality…”

And, yes, this is still about reliability. The expensive thing for AI agents isn’t compute; it’s failure. That’s why Flo keeps adding scaffolding:

“We call it the rule engine… a verifier wrapping every step of your agent… you can define your rule policy… soft rule: try up to three times then proceed; hard rule: must be true or don’t do anything… It all compounds.”

That’s it! See you in your inboxes on Sunday.