AI and agency

Last week I went on an old-fashioned tweetstorm , reacting to this actually very insightful and smart thread by Atlassian's Avani Prabhakar about that company's AI adoption journey.

Short version: Atlassian has studied a few of their highest-performing, most-AI-adopting "frontier teams" and found a few patterns that seem to contribute to these teams seeing the most benefits from AI. What they've found is that clarity, psychological safety, and focusing on outcomes, not outputs were key ingredients in teams being more successful with AI.

Of course, these have been key ingredients for successful teams even before AI. A decade ago, Google's Project Oxygen and Project Aristotle studies concluded that teams with high agency and psychological safety performed better than ones where team members were asked to just buckle down, do the work, and not ask questions.

My thread responding to Avani's — after recapping the whole history of Agile — offered these ideas in response:

1. Engaged, empowered teams will tend to perform better under any circumstances (i.e. the Project Aristotle conclusion)
2. Engaged teams who like to experiment with new tools will find the useful edges of those tools faster, and be more productive with them, simply for love of the game
3. Bottom-up tool/technique adoption will almost always go better than top-down mandates; in fact, where bottom-up adoption may massively improve outcomes, top-down mandates will often do the opposite.
4. Companies known for great product/engineering culture, like Atlassian, are gonna have a smoother, more successful AI rollout than others.

I'm in some Slack and Discord communities with folks who work in engineering orgs that have mandated AI adoption, including some that have tried token leaderboards — at least until people were caught "tokenmaxxing" on useless tasks to artificially drive up their numbers . (See also Goodhart's Law: "Once a measure becomes a target, it ceases to be useful as a measure.")

​ John Cutler wrote about this tension around AI adoption and agency today in his latest newsletter:

“Many leadership teams are engaging in what amounts to collective gaslighting of their teams. They treat resistance to AI as a personal failing, a lack of curiosity, motivation, or willingness to adapt, rather than a rational response to a changing environment. … But as Bandura put it, “Personal agency operates within a broad network of sociostructural influences.” People are both producers and products of their environment.”

As both a technologist and a web agency owner, lately I've been trying to reconcile a few seemingly contradictory observations:

Many people genuinely love using AI. Whether or not their AI answers or documents are any good, or good for society, it's clear that folks are really having fun yapping back and forth with a machine intelligence and ending up with what looks like work product. I've had clients send me AI-generated 'wireframes' that are more like full mockups , and their enthusiasm for having "created" something so "nice" so easily is palpable. This is the aspect of AI that Craig Mod describes when he says he's "software bonkers." ​

Nearly everyone hates feeling forced to use AI. This runs the gamut from employees chafing under new performance review criteria and token-use monitoring, to gamers switching to Linux because seeing a Copilot icon appear in the Windows taskbar was their last straw. When it's your idea (or it at least feels like your idea) to use a prompt, it's good. When a prompt box shows up in every app with toast messages begging you to prompt, it's bad.

Everyone hates 'slop', including the people producing slop. Just based on how universal the anti-slop sentiment is , compared to the rate at which slop is being produced, it's obvious that many people who are creating slop also say they hate slop. This would imply that some people don't think slop is slop when it's theirs, which is to say, they expect others to put in more effort but want grace if they take a shortcut. Like the agency thing, this isn't new or surprising; AI just makes it more obvious.

AI is massively increasing the amount of stuff thrown at us, but the outcomes and value aren't there yet. I've seen multiple surveys and anecdotes on this—folks are having fun making apps and services with the help of coding agents and tools like Lovable, some of them thinking they've struck gold now that their amazing ideas are no longer gatekept by technical or design skill. Then their project launches to zero users . 🦗🦗🦗🦗🦗🦗

Taking all of these together — Atlassian's high-agency, high-success AI teams, people enjoying producing AI output but hating receiving it, the huge number of AI-generated projects with zero audience — it seems like AI is great at lighting up a room for one person (its user), but is still quite bad at bringing people or teams together. And that is where the greatest value is found.

This has been on my mind a lot because, while (like other studios) I help build websites, which a lot of people see as brochure-like artifacts that Claude or other AIs can easily crank out, I'm really focused on the systems behind websites that enable teams to run them sustainably.

I wrote recently about the apparent tension between AI coding agents and CMSes , which I think is really about whether people do or don't grok how teams collaborate through and around systems, and how that collaboration results in web pages for their users.

Ironically, an AI coding agent can broadly give teams (especially dev teams) more agency overall, especially compared to walled-garden platforms like Webflow or Wix. But that particular agency isn't evenly distributed; it relies on everyone feeling OK having to ask Claude for something that could be done as fast or faster in a text field in a CMS.