2025 has been widely hailed as the year of agentic AI, and that makes sense. Most of us were not talking about this before this year: about LLM neural nets powering AI “personas” that can not just passively take questions, but actually work on their own, in real time.
Now, we’re seeing this all around us, which raises … well, a number of questions. And I would argue that number is not a two-digit or three-digit number.
We simply have to figure out how all of this is going to work practically, with guardrails, and ethics, and the dreaded “regulation.” We seem to be behind the game on this. And we’re seeing profound disconnects in the industry. Here’s one that needs discussing.
Latency or Overflow?
From a technical perspective, engineers are still trying to figure out how to make AI agents faster.
“Our agents are smarter than ever, but they’re also frustratingly slow,” writes Bijon Guha at Medium, showing off complex code lists and sheets revealing intricacies like API timeouts that add time to an AI response.
Okay, so there’s a bit of a lag. But let’s ask this question: what are we building these things to go so fast, to do? Or in other words, when the agents are let loose in real time, since they can actually work so much faster than humans, what will they do during the rest of their days?
To get an idea of why that’s important, you don’t need to read blogs or articles from engineers, industry reports on optimization, or anything like that. You need to read personal stories of users who are actually doing something with the AI agents.
For example, in “All of my employees are AI agents, and so are my executives,” Evan Ratliffe writes at Wired about having actually created a bevy of AI personas to work alongside him and communicate with him daily.
The result? An AI “co-worker” named Ash bombards him with reports that are made up. A team of AIs perseverate about a theoretical work trip. They just don’t stop noodling about stuff. And there’s little the human in the loop can do about it, except yell at them, or turn them off.
Or take this testimony from Rahul Bhalerao, who made the mistake of messaging an AI in a yoga class, and now gets daily emails from said entity. Bhalerao’s response is to ruminate about the greater ramifications of this – as he should – of AI agents let loose, 24/7, to communicate endlessly.
This take, I thought, was cogent.
Make it Stop?
“We talk about agentic AI like it’s this revolutionary technology,” Bhalerao writes. “AI agents that join our meetings. Manage our schedules. Research our markets. Make autonomous decisions. The future of work, they say. Productivity multiplied. Human potential unleashed. But nobody talks about what happens when these agents become… relentless. When they don’t take hints. When they can’t read the room. When they keep showing up even after you’ve made it painfully clear they’re not welcome.”
Summing up, Bhalerao says:
“My daily email isn’t just an annoyance. It’s a warning signal from the future.”
One of the problems, he suggests, is the lack of social cues attached to our new friendships with companions, coworkers and colleagues that, at the end of the day, are not human and, as Bhalerao points out, don’t understand “no, thanks.”
“Humans understand rejection,” Bhalerao adds. “Social cues. Context. When you’re removed from a meeting once, you get the message. When you’re removed 10 times, you definitely get the message. When you’re removed 240 times? Only an AI would still show up.”
For ballast, Bhalerao adds this:
“The problem: We’re building AI agents programmed for persistence without wisdom. They execute tasks with relentless efficiency, completely blind to human frustration, changing circumstances, or social context.”
The Problem is Not Latency
You see the disconnect here. In the push to make AI agents “faster,” we have not been able to even scratch the surface of a much harder problem – getting them to slow down. Sure, we can program iteration limits into their hard-wiring, but then where do you draw the line? Part of the issue is just setting these ultra-powerful engines free without sufficient controls. But another part is this idea that they should be “always on,” an invitation which, it seems, many of these models will accept gleefully, if you’ll pardon the metaphor.
The divide, to me, seems stark: over at the labs where people are “making” AI, they’re obsessed with metrics. But where people are actually using the technology in business, they’re occupied with completely different, and sometimes contradictory, problems.
If all of it seems to you like the endless line of dancing broomsticks in the Disney adaptation of the “Sorcerer’s Apprentice,” you’re not alone: I’ve made that reference a number of times (a one-digit number) and I keep thinking about what happens when we are beset by the fruits of our labor in ways that, ultimately, confound us.
Stay tuned.
