#034 - Agent double-0 nothing
A look into agentic AI's hazy present. And possible future improvements. With some ideas from algorithmic trading.
You're reading Complex Machinery, a newsletter about risk, AI, and related topics. (You can also subscribe to get this newsletter in your inbox.)
(This one turned out much longer than I'd expected. The first half is a summary of where agentic AI is today. If you're in a hurry for the "how does it get better?" insights, you can skip straight to the "Bot lanes and middlemen" segment.)
I've been meaning to write about agentic AI since the term first landed late last year. AI Stuff™ kept happening, though, so I kept pushing it off. Today I figured I should get this out before the agentic hype wave collapses and we are presented with the next shiny bauble of distraction.
As I promised last time, this will be a tale of hype, hope, interfaces, and intermediaries. And quotes. Lots of quotes. Agentic AI ties together points I've made in past newsletters.
It was all a dream
In Rebranding Data I noted that the data field changes its name every few years to avoid a much-needed market correction. I penned that piece before the AI field had renamed itself to … AI all over again (as people often drop the gen- prefix). It still holds.
The arcs within genAI of special concern. These are shorter iterations, a new toy arriving just as reality starts to catch up with the previous one. That, and genAI has led to a change in theme in the data field – from "analyze data for fun and profit" to "replace people." I get that technology is automation, and that automation eats work, yes. But LLM-mania has pushed this into overdrive. Companies dream of (supposedly, intelligent) machines freeing up large swathes of the org chart.
Which brings us to the latest flavor of genAI, called "agentic AI." It's not entirely clear what this all means, but it seems to involve AI bots – the eponymous agents – tackling a wide variety of real-world tasks on your behalf. Plan your travel, buying movie tickets, placing orders for groceries. That kind of thing. But, again, the definition is hazy.
(The lesson we keep missing is that ill-defined terms make great hiding places for charlatans. So when an entire market is hyped over something no one can explain … well … keep an eye on your wallet.)
Hmmm.
In just a couple of years genAI chatbots have offered terrible cooking advice, misinformed customers, mocked their creators, defended the un-defendable, and failed Ethics 101. So it makes perfect sense that we would [checks notes] give them even more authority and less supervision.
Right?
Right.
What?
The future's tense
I'm trying to keep an open mind here. But notice that most of the agentic AI discussion is set in the future tense. Something I warned about few months ago:
AI vendors keep getting tripped up on their Fake It Till You Make It routines, so they've changed tactics. They're moving away from concrete promises for today, and – taking a page from the cult playbook – declaring future dates for when things will pay off. "It's gonna be amazing in just a couple more years. You will be rewarded for your belief. Trust me. And also, pay me." This gives them extra time to turn their bold-yet-empty proclamations into something real. Or to wriggle out of the promises if things don't pan out.
(You know how cults always find some excuse when the big date passes without fanfare? With AI, they'll just rename the field and start over. Again. You heard it here first.)
That, in turn, pairs well with something I explained in January:
[T]he people buying and selling AI live in different times.
Tech companies, their fans, and their investors all live in the future. The companies' pitches focus on the possibilities of a product that has yet to be built. "We don't have anything yet, but we will, so we'd like our valuation to reflect that." Their investors dig this because they get paid years down the road, when the company goes public or gets acquired. And the fans, they see quirks and glitches as a small price to pay for cool functionality. All three groups are dead-set on deriving benefit Sometime Later™ .
Compare that to the consumers and businesses buying that technology. These people live in the present and they need things to work Right Now™. Future-dwelling tech companies forget this when they travel to present-day land to sell their goods. They use present-day terms to tout their products' future benefits, and they're confused when buyers (who live in the present) complain about malfunctions. "I mean sure it didn't catch the bad guys / summarize your headlines / enhance your search, but … imagine what it will be able to do later! Isn't that great??"
It's hard to not see agentic AI as the latest flavor of Tomorrow's GenAI, Sold To You Today. Which wouldn't be a problem, except that Yesterday's GenAI doesn't work today. GenAI comes off like a salesman who is in a race against time. Constantly looking over its shoulder, staying out of reality's reach.
Executive ambition
Buyers are already sold on the future idea, though. They very much want to like genAI's present, except that the use cases they've surfaced don't strike me as agent-worthy.
Take this Bloomberg article from December. It's about executives touting their agentic AI plans but it reads more like a wish list. One part in particular stands out:
At consulting firm McKinsey & Co., an AI agent now handles the tedium of client onboarding. It coordinates paperwork, shares relevant contact details, affirms the scope of the project — and runs everything by the firm’s legal, risk, finance, staffing and other departments to get their signoffs.
“That used to be an absolute spaghetti bowl of email threads between all the different functions,” said Rodney Zemmel, who leads McKinsey’s digital practice and the firm’s own AI transformation. In the past, onboarding required tens of hours per new client. Now, “an agent basically does all that chasing for you,” and completes the process in roughly 30% of the time. It sends emails and follows up to wrangle whatever information it needs to move projects forward. The final product is then reviewed and approved by a human.
It's not clear why a shop like McKinsey would trust client on-boarding to the AI version of a customer service phone tree. (As a consultant myself, I'm surprised that the on-boarding is that messy in the first place. Maybe that's just me.)
Also, right as I thought "isn't this more of a job for a rules engine?" I got to the next paragraph:
“It works in this case because it's a complicated set of tasks, but actually a fairly standardized and routinized one, without too much judgment involved,” Zemmel said.
This … does not require AI, then. Not genAI, not ML/AI, none of that. If you have established firm business rules to handle "fairly standardized and routinized" tasks, then Plain Old Software will do just fine. Bonus: it will be cheaper to run and easier to support, by virtue of being deterministic in operation.
In a separate article, Ford's agentic AI plans gave me a similar vibe:
WSJ: This seems like you’re giving the AI an input and getting an output rather than an autonomous agent that completes a series of tasks. When does it become agentic?
Goodman: When we string them together: so the agent might be rendering and then creating a 3D model and then doing a stress analysis on it. That’s where I view this as agentic.
I wasn't present during this interview to ask clarifying questions. So maybe I'm missing something. But as written, this seems like another job for Plain Old Software.
And that's the scary part: it's not just the vendors who are talking about how amazing agentic AI will be. The customers are doing it, too. They've memorized the slogan and they're chanting it to their peers in some kind of corporate hypnotism act. This has ceased to be about AI – Analyzing Data For Fun And Profit – and swerved hard into Sell The Dream territory.
I understand why the companies selling genAI are pushing hard on this. They have a financial interest in keeping this dream alive. But as for the companies buying genAI who are playing the In The Future game, I'm stumped.
AI in general, and genAI in particular, is still an immature technology that is inherently probabilistic. Leaning on genAI – and its inner wild animal, The Random™ – when Plain Old Software will do, that seems folly to me. Folly, and a way to take on tons of needless risk exposure with very little return on the horizon.
Bot lanes and middlemen
I'm not saying agentic AI won't work out. I'm saying that agentic AI will require some changes in order to reach its full potential.
The obvious changes would be to dial back on the hype (to focus on what agents can actually do) and then work on improving the bots' capabilities (so they can do more). The latter will require luck as well as research effort, but it's within reason.
Two additional, more-subtle changes will have an outsized impact on agentic AI's future:
1/ Creating AI-specific interfaces. When computers came to Wall Street, companies didn't put humanoid robots in a trading pit and train them in the words and hand gestures of their human counterparts. Instead, those bots plugged into the exchange's data feeds, where they could inhale a flood of market prices and issue orders. The robots spoke like robots, to robots, and as such they communicated at robot speeds.
(If trading's not your style, there's an analog in the Star Wars universe. Notice how the droids can plug into a central computer to open a door or get information. Compare this to the way everyone else has to use manual controls or issue voice commands.)
Today's agents are taking the opposite route. They work by clicking through websites, a process that is clumsy and error-prone because those interfaces are built with people in mind. The agentic AI sales pitch is to hand work to the bots, right? So why are we making the bots use the human interface? It's a better idea to build lanes just for them.
(We can make a similar statement about autonomous vehicles. But that's a story for another day.)
This isn't idle dreaming on my part, either. I spent my early career building those kinds of backend, machine-to-machine systems. I acknowledge that it takes a lot of work and planning. But a good deal of the groundwork has already been done. Today's websites and mobile apps are usually tiered affairs – the clickable, user-facing layer issues API calls to backend systems that handle the business logic. One could create a similar, agent-facing layer to run separately from the human-facing UI/UX. And it would reduce the scope of errors the bot could commit, because it's not looking for a button that's labeled "Submit." Or was that "Place order?" Or does this site use a cutesy "Let's go!"?
That makes things better for the robots. Now let's talk about making them better for humans:
2/ Providing intermediaries to protect end-users. When people test AI agents for tasks like online shopping, a recurring theme is for the humans to take the helm – grab control of the on-screen browser the agent is using – in order to enter sensitive details like credit card numbers or double-check matters before completing a purchase.
Longtime readers already know my take on middlemen, but this is a place where middlemen could shine. Imagine a third party, inserted between the agent and the merchant, to hold those credit card details and flag odd purchases.
Once again, this isn't idle dreaming on my part. Such middlemen already exist. The iOS App Store and Google Play Store have your credit card on-file. When you make an in-app purchase, the app developers get the money – through Apple or Google, not directly from you – yet they never see your payment details. And speaking of credit cards, big payment networks like MasterCard and Visa run their own fraud detection programs on top of those used by your credit card issuer.
LLM providers are well-placed to fill this role of the intermediary in the agentic world. You already have a credit card on-file with OpenAI, Anthropic, and the rest to pay for chatbot usage. What if they could extend that to purchases your agent makes? Add a simple system to note that, say, your proposed total from this restaurant looks suspiciously high. (These enhancements would require the bot-specific lanes I mentioned earlier. But those need to happen anyway.)
The app stores and payment networks serve to identify, mitigate, and – most importantly – absorb consumer risks. Early-day e-commerce customers felt safer entering their payment details into websites because they knew their credit card issuer would cover fraudulent charges. Today we think little of buying in-app items because we know Apple or Google can step in if the app-maker pulls a fast one.
Consider what these intermediaries did for consumer confidence, and therefore, for adoption of online shopping and in-app purchases. Our world of e-commerce, rideshare, and food delivery apps is possible only because someone else said "I'll saddle some of the risk in order to grow this space."
If agentic AI providers are serious, they'll find a way to do the same.
Have my people talk to your people
To close out, let's consider one more intermediary for agentic AI. This one would look very much like a trading exchange – sitting between the parties, settling rules, and handling disputes.
I'll set the stage by raising my (recurring) point that the rise of ML/AI parallels the rise of machines on Wall Street. I keep thinking about something I said in newsletters #016 and #017:
If this sounds like the bot-on-bot world of algorithmic ("electronic," "computerized") trading, that is precisely where my mind is going. Or, better put: that's where my mind naturally gravitates when I think about AI. As I've noted elsewhere, the story of computers moving into financial markets – from the introduction of REG-NMS, to runaway trades and flash crashes, and everything in-between – will tell us a lot about where our AI-enabled world is heading.
That "bot-on-bot action" I mentioned? Agentic AI brings us one step closer to this becoming a reality. And that ties to the point I raised earlier about machine-specific interfaces: it won't be long before consumer-agents stop clicking on merchant websites and start talking to merchant-agents.
Hmm.
We know what a flash crash looks like on Wall Street. We also know the role the exchanges play in handling crashes and runaway markets – everything from busting trades to establishing circuit breakers.
What will it look like when agent-based grocery bots have a bad interaction, or when fraud detection bots get stuck in a loop, or when credit-scoring systems collide? And who will step into the middle to set things right?
We'll find out. And if the answer is "agent providers will heap all of the risk onto consumers," that will tell us a lot about how far agentic AI will go.
In other news …
- WSJ tech reporter Joanna Stern explains her love-hate relationship with AI-based search. It's mostly love. (WSJ)
- Do you think Meta
piratedsampled your work to build its AI? You can check on that. Sort of. (The Atlantic) - Meta continues its demolition of content-moderation teams. This time, in Spain. (Le Monde 🇫🇷)
- People are already using genAI bots to simulate difficult work conversations. Is it a surprise that they'd also use them to profile romantic interests? (FT)
- Given … "recent events" … the rest of the world sees US-based tech companies through a new lens. And now they're looking for alternatives. (MIT Technology Review)
- AI startup Perplexity denies that it's in financial straits. But it also declares that it won't IPO till at least 2028. Hmm. (TechCrunch)
- Chatbots actually make lonely people … lonelier. (Gizmodo)
- People are using genAI to create images in the spirit of Studio Ghibli's animation. Not everyone is so amused. (CNN)
- Because who wouldn't want a 24/7 electronic eye watching over them in a mental health ward? (The Guardian)
- It's bad enough when The Algorithm™ downvotes your video or social media post. That's nothing compared to when it terminates your employment. (New York Times)
The wrap-up
This was an issue of Complex Machinery.
Reading online? You can subscribe to get this newsletter in your inbox every time it is published.
Who’s behind Complex Machinery? I'm Q McCallum. I think a lot about AI and risk, which I write about here.
Disclaimer: This newsletter does not constitute professional advice.