Plainly speaking, if you think your AI agent is going to replace teammates, you never valued your teammates to begin with. Reading John Cutler's linkedin post about what makes a team, has me thinking about the proliferation of so-called agentic management frameworks. Agentic-replacing employees logic only works if you've already decided that the only thing that matters about your teammates is what shows up in their outputs.
Some of my favorite things about working in places with rad coworkers are the rad coworkers themselves. Not just being able to vent to them about random facts, hearing about their home renovation, or grabbing a drink after work sometimes, but the things I learned from them and the good feelings you get from being trusted by a stranger who maybe sees potential in you, encourages you, offers feedback at a crucial moment.
Even outside of the personal relationships, as a team sports guy who also coaches high school sports, I just appreciate coworkers the way you would a teammate. It's fitting that most people who lead [anything] these days never played a competitive sport or attempted music at any point. The colleague who flags the thing that's about to break before anyone else sees it, the one who trains the new hire without being asked, the one whose memory of a decision the org made seven years ago keeps you from repeating the mistake are the people you need to preserve.
I read John's post mostly as a lot of managers do, people sitting around wondering who is slacking off so they can fire them, and others worried about how to divide the credit. As a fan of sports, I am obsessed with analytics and have been since the 90s when it was really hard to get data besides Baseball Weekly. The stat sheet simply cannot measure a lot of intangibles well for you. There's an old article from the New York Times Magazine that talked about former NBA player Shane Battier and how he was the quintessential "no-stats all-star." Looking at his numbers it would not seem apparent that he's all that useful, and yet people who know ball including his peers in the NBA appreciated the intangibles of his contributions. Draymond Green of the Golden State Warriors is considered a future Hall of Famer for being a glue contributor to a dynasty despite never being known for his scoring.
The same anxiety driving managers to wonder who's slacking off is what's driving the agentic management gold rush. If you measure people solely on outputs, an agent that produces outputs looks like a viable substitute. It'll make you think you can cull headcount, brag to your boss about productivity being up and just hope to G-d that no one checks the work. The idea that we can create onboarding protocols for software is cosplay.
None of this is any different than the guy who declares his backyard a sovereign nation and starts issuing passports. The Principality of Sealand has a flag, a national anthem, royal titles, and a stamp collection. None of that makes it a country. Building HR functions for stochastic text generators is somehow even more embarrassing than making passports and currency for your backyard. At least that's kind of fun? The most offensive part of all of this cosplay is the fact that we already don't treat our human employees as well as we can, and there are people with a straight face asking us to consider agentic welfare. It's embarrassing.¹
Having managed an agentic staff for a while now, I can say the best part of this revolution is having stuff that used to be too expensive to build now cheap enough to put together and offer real value. One example is oregontennis.org, a site I’ve owned for years and finally developed the state’s first high school tennis team power ranking this year because it took me hours rather than weeks to aggregate the data, build it, and maintain it. Thousands of people use it, and it costs me almost nothing to keep online.
Agents are useful, and agents are also tools, in the way a printer is a tool. Nobody thinks the printer is a colleague or asks whether the printer is on the team; we're not building parallel HR functions for our printers. Someone still has to know what to send to the printer, when to send it, how much to print, when the printer is making errors, when to replace the printer toner (which apparently is now IP?), and when the job is one that the printer cannot handle, no matter how you configure it. That work stays when you buy a better printer. A better printer makes the work more skilled.
Agents work the same way, and the best way to manage them is knowing how to do the thing you're trying to do faster before you delegate the task. Unsurprisingly, this is true of managing people too. How can you truly judge the quality of a work product if you don't understand what goes into it? I'm not suggesting you have to know everything about every part of the roles within your purview, that's why you hire people to begin with. The difference between evaluating and vetting good human candidates based on past experience, training and learning, is that even if you can't evaluate every aspect of what they've learned, at least you can relate to what that journey looks like.
The judgment and evaluation pieces of the work cannot be delegated no matter how we want to believe that doing so will make things easier. I'm not talking about generalized rubrics or giving people access to stuff they're eligible for in a blanket way, I'm talking about the legions of edge cases that will proliferate and have real impact on people's lives, as we delegate human decision-making to tools, allowing them to intercede into processes that they were never intended to be part of. Nobody wants to talk about the judgment layer enough, because there's no money in slowing down right now. At some point, you're going to need to and the ways to do it are just going to cost you money.
You can have agents and you can have a workforce that gives a damn. The frameworks proliferating right now are not the way to get both, because they presuppose that we've wrung everything we can out of disposable people and the machines can take it from here. Maybe they can. But the cost of being wrong, like all of the decisions of the platform economy, will be paid for downstream and not by the people making them.
P.S. Speaking of things I built earlier this year... Occupant publishes three daily-updated indices on AI compute economics. Building consumer-price-style measures for an ephemeral product is noisy work, but consider it one piece of a much larger conversation we should be having about AI spend, token churn, and the downstream impact on everyday life as this software embeds itself into everything.