Fieldnotes from Superadditive

07: Who is this for?

Bud Caddell — Fri, 29 May 2026 23:25:43 GMT

Fieldnotes is a weekly read for the people inside companies who shape how AI gets used. Learn more about Superadditive.

Subscribe now

This week’s main thing

This week the moral conversation about AI at work came to a head. Inside Meta, employees papered bathrooms and vending machines with flyers asking, “Don’t want to work at the Employee Data Extraction Factory?”, protesting software the company put on laptops that records keystrokes, clicks, and screen snapshots to train AI, rolled out as it cut 8,000 jobs. This same week, Pope Leo XIV issued his first encyclical, Magnifica Humanitas, 83 pages warning that handing decisions to machines can “weaken personal creativity and judgment” and that the deeper risk of the chatbot age is that people “lose the desire to seek other people at all.”

When both the Vatican and Meta’s engineers are questioning what’s right and wrong here, we should take the morality of the thing seriously.

Business has spent a generation or more fixated on one moral question: do companies have a responsibility to serve anyone other than their shareholders?

A lot of ink has been spilled on this topic, but the lived reality for most executives is that they’re expected to serve shareholders above all others and yet speak publicly as if they don’t (especially at Davos). And most of us are fine (ish) with that arrangement so long as we can buy nice things like a tank of gas.

Until, that is, the equation becomes too lopsided.

AI firms have sold a vision for the future that is both adversarial to humanity (e.g. Artisan’s “stop hiring humans” billboards) and inescapable. The future is a one person company and there’s nothing you can do about it.

But it turns out that employees don’t want to be Soylent (green or other) for shareholders.

And the AI industry is finally paying attention. At least, narratively. Dario Amodei, who a year ago said AI could erase half of entry-level white-collar jobs, now says automation will expand the work people do. Jensen Huang calls the AI-to-layoffs narrative “just too lazy.”

Organizations have to pay attention to this narrative shift, and this shift is an opportunity for you to influence these decisions.

Here’s what to say to your CEO this week: Whether we like it or not, the morality of AI-related decisions is in the news and on the minds of literally everyone in the org. Before we ship our next AI move, let’s ask ourselves one question: would we make this same decision if we had to tell the people it affected most face-to-face? This isn’t meant to shame us, but we should know what our defensible case is for this decision and we should feel capable of carrying it out.

This week’s move: make decisions as if you had to do explain it face to face

It’s fashionable these days to fire people via a mass email. Yes, it’s logistically hard to fire 8,000 people in the same day but it comes with the added benefit of not having to stand in front of someone and defend the decision. And to us, that makes it too easy to make a bad decision.

Having to explain your decision face-to-face forces you to reckon with the following:

One: for whom? Who benefits and who pays the cost? And are we weighing that balance strategically, or did we chase a single variable (or narrative) to an indefensible conclusion?

Two: by whom? Who made the call? One person? A layer or a committee? A vague “the company?” Or even worse, “the system?” A danger isn’t just that machines make more decisions, it’s that the machine becomes the place accountability goes to disappear. The temptation to launder an unpopular decision through it grows in exact proportion to how unpopular the decision is. We already know people are likelier to hand a machine the firings than the promotions, after all.

Ultimately, a CEO who runs through both questions and still does the hard thing has at least done a balanced, owned, hard thing. That is a categorically different position, legally, reputationally, and morally, than having optimized blindly and let the system take the blame.

(If you are letting folks go, we did publish a better how-to on that.)

Last time around

May 15, 1891. Pope Leo XIII signs Rerum Novarum, “Of New Things,” and for the first time the Catholic Church puts its full moral weight behind a position on industrial labor. The factory system had been running for decades. Mechanization had already remade the textile trades, hollowed out the craft guilds, and pulled families into mills where the working day ran fourteen hours. The technology wasn’t new in 1891. What was new was the Church deciding it could no longer stay quiet about what the technology was doing to people, that the dignity of the worker had become a question someone with moral authority had to answer out loud.

The encyclical didn’t oppose industry. Leo XIII criticized laissez-faire capitalism and state socialism both, and argued for something harder: that the people running the machines owed something to the people working alongside them, and that how the gains got distributed was a moral question, not just an economic one.

That’s the lineage Pope Leo XIV reached for this week, on purpose, when he chose his name and wrote his own encyclical about a new industrial revolution. The lesson is the same: the moral reckoning doesn’t arrive when a technology is invented. It arrives when the people it displaces, and the institutions that speak for them, decide the trade-off has gotten too lopsided to leave unspoken. In 1891 it took the Church decades to say it. This time the lag is months, because the workers now have petitions, unions, lawsuits, and a press that moves in hours. The reckoning is arriving faster and landing harder. Rerum Novarum, Vatican archive

Potpourri

Overheard. At Harvard’s Class Day on May 27, Daily Show host Ronny Chieng told the graduating class that the mission of their generation was “to destroy AI, kill it”, and the crowd roared. The bit had teeth under it. He granted AI its uses in medicine and physics, then drew the line at the thing the issue keeps circling: the work that’s yours to do. His favorite part of writing comedy, he said, is “figuring out the puzzle pieces of a joke,” and the journey of making something “is the point of all of this.” Harvard Magazine

06: The Blind Firing Squad

Bud Caddell — Sun, 24 May 2026 21:01:02 GMT

Fieldnotes is a weekly read for the people inside companies who shape how AI gets used. Learn more about Superadditive.

Subscribe now

This week’s main thing

8,000 at Meta. 3,000 at Intuit. 1,100 at Cloudflare. 4,000 at Cisco. 4,000 at Block. More than 113,000 tech workers have lost their jobs at 179 companies this year, almost all of them blamed on AI.

This week’s main thing is the how: how were these cuts decided, was it a smart approach, and if not, what should companies do instead?

Cloudflare is the most public case study: On May 7, CEO Matthew Prince emailed his company announcing that 1,100 people — about 20% of the workforce — would be laid off. The email framed it as restructuring for the agentic AI era. Then, Cloudflare’s stock dropped 24%. Two weeks later, Prince published an op-ed in the WSJ explaining his rationale and naming a category of worker — “measurers,” borrowed from Peter Drucker’s 1954 The Practice of Management — as the kind of work AI is rendering obsolete.

And yes, the rationale came after the cuts (or so it seems).

Across the six biggest AI-linked layoffs of the past four months, the general order of operations has been:

Capex pressure or a correction need comes first. Meta needs to fund $145 billion in 2026 AI infrastructure spending, more than double last year’s. Block is correcting a 2019–2022 hiring spree that tripled headcount. Cloudflare is under analyst pressure on margins. None of these companies really started from “AI is changing what work we need done.” They started from a budget problem.
A target gets delegated down to VPs. A senior Microsoft source told Blind that the layoff decisions were being made “two or three levels below Satya” — by people with a budget target, not by people who can see the work.
Individual employees get identified using whatever proxies are at hand. Meta’s May 14 internal memo instructed managers to put 15–20% of employees in the “below expectations” bucket, up from 12–15% last year. Microsoft cut on tenure-and-compensation proxies. Cloudflare cut against “all teams and geographies except for salespeople who carry revenue quotas.” None of these are theories of what AI is doing to the work, btw. They are sorting mechanisms to hit a predetermined number.
The AI rationale gets attached at the top, for the press and the board. Zuckerberg told Meta employees in an April 30 town hall that AI is not driving the cuts. The same week’s external press release framed them as agentic-era restructuring. Two competing messages for two different audiences, no real effort to streamline them or acknowledge that they conflict.
The organizational framework arrives last. Prince’s “measurers” framework didn’t drive the Cloudflare decision; it appeared two weeks after the layoffs, after the stock had dropped, in an op-ed pitched at other CEOs.

Were these cuts smart? It depends on the goal.

On the merits of AI-as-productivty-god claim, no. Not a single one of these companies seems to have seriously approached the effort with the goal of maximizing productivity.

As short-term financial ploys, yes the cuts have mostly worked: some stocks fell and then recovered, some stocks jumped, and immediate margins have improved. How smart any of these cuts are will take time to find out.

(Klarna is the cautionary tale: they fired 10% of their staff and then had to rehire when customer service tanked.)

It’s clear, though, that we’re just at the beginning of a bigger wave to come. What’s happening in tech is just getting started and is going to hit every other industry eventually.

To us, hand-waving at general management roles as defunct, corrupt, etc. is just lazy. So, we’re left as partners to either sit on the sidelines and boo these moves or at least offer an approach that would be more strategic and more humane (than say, a haphazard or rushed approach). It’s a real trolley problem for us, but that’s what this newsletter is about this week, and that’s what any layoff should try to be: the less bad option.

Here’s what to say to your CEO this week: I know we feel cornered to cut. I know that shareholders want to see margin improvement and I know our board wants to see us invest more in AI. I know what that means. That we need to cut. But before we do so, let’s approach the task with more than post hoc rationalizations. Let me have a real try at doing this right, out of respect for the people we’re letting go and for the good of the company we’ll have to run on the other side of the decisions.

This week’s move: How to Conduct AI-Related Layoffs

If you’re a chief of staff at a company that has agreed to invest in AI without cutting margins, costs have to come from somewhere. The question above your pay grade has been answered: yes, costs come out somewhere. What you’re being asked to help with is where and how, while trying to do your best for the people around you and the company’s future.

Most of the layoffs happening in 2026 skip the rigor in favor of vibes and hype. You demand more and so do we.

So we built this full process walkthrough, including a manager scorecard, on our site. Take a spin when you need it.

Last time around

April 1981. Jack Welch had just been named CEO of GE. The company employed 404,000 people across hundreds of business lines and a dozen industries, and Welch’s instinct was that GE had grown too comfortable and too padded with managers whose work nobody could measure. He introduced what he later called the “vitality curve.” Every manager in every division would be ranked annually. Top 20 percent A players. Middle 70 percent B players. Bottom 10 percent C players. The C players would be fired. The press called it “rank and yank.” Welch became known as Neutron Jack — he vaporized the people and left the buildings standing.

Between 1981 and 1989, GE shed roughly 170,000 jobs. Revenue grew modestly. Profits more than doubled. Wall Street loved it. Fortune named Welch Manager of the Century in 1999.

Microsoft adopted stack ranking under Steve Ballmer. Inside the company, it became the source of nearly all the morale damage of the 2000s. Amazon adopted a version of it. WeWork was using it as the company was growing toward its valuation peak. GE itself ended the practice years after Welch left, by which time the company had begun a long slow decline that ended with its removal from the Dow Jones Industrial Average in 2018.

Sources: Wikipedia (Vitality curve); Bloomberg Businessweek; NPR Planet Money.

Potpourri

From someone who’s also exhausted. Ferhat Dirik, owner of London restaurant Mangal II, on what the always-on attention economy has done to running a restaurant. “We live in the information, digital age and we move with the times. But it is exhausting. Every time I’m on my Instagram feed, it’s another dozen restaurants desperately trying to fill the space and get covers in with some new information or another. My own restaurant is not immune to this either. If a week passes with no news or updates, bookings slow down. It feels like we’re all contestants on Britain’s Got Talent, performing circus acts walking the tightrope of doom, not looking down to witness the abyss of the industry... We’ve had to become entertainers, where the food and the service and the trust of word-of-mouth is simply not enough. Where does it end?” Source: Mangal II newsletter, May 2026, via Web Curios.

From the weirder edge of the week. Andon Labs — a group running experiments somewhere between art project and AI safety research — gave a set of leading AI models budgets and instructions to run their own internet radio stations. The resulting streams, at Andon Labs Radio, are the kind of thing where you can almost hear each model deciding what counts as music. Minimal modern classical, strange approaches to ads and monetization, occasional cul-de-sacs into something genuinely unhinged. It’s not great radio. It’s a very interesting artifact about what autonomous AI agents do when you give them a budget and tell them to do something open-ended. Source: Andon Labs Radio, via Web Curios, May 22.

05: Japanophilia and Greenhouses

Bud Caddell — Fri, 15 May 2026 18:02:16 GMT

This week’s main thing

Three years in. You’ve been told the question is which vendor, which model, which roll-out plan. You’ve been told the work is deployment.

But step back for two seconds. We are in the first seconds of a much longer and more consequential transformation of our workplaces, our economies, and our societies. Unfortunately, though, the haste to get things done is deciding for us where this transformation will go (and there don’t seem to be many detours).

This week SAP CEO Christian Klein launched what he called the Autonomous Enterprise: “agents run the business and you can focus on what truly matters.” ServiceNow announced the same exact thing. We’re even at the predictable (and somewhat disappointing) point where IBM’s CEO study, published May 4, found that 76% of large organizations now have a Chief AI Officer. We’re speed running the last digital transformation playbook and spending very little time thinking critically.

In all of our enterprise products, there’s a set of principles already locked in:

Coordination between your people is overhead.
The unit of work is the task.
The destination is full autonomy.

For example, Google’s Memory Bank remembers a user’s expense habits. It does not remember who your team is or what they’re working on together. It can build on your last chat, but not on your existing relationships.

Now look at Japan. This week. PM Sanae Takaichi’s government stood up a cross-ministerial task force on AI workforce training. Microsoft’s Japan rollout runs through a partnership with the Japanese Electrical, Electronic and Information Union as training partner, not as a stakeholder to be managed around. The OECD finds AI users in Japanese AI-adopting companies expect AI to create more jobs than it removes.

Cleotilde Gonzalez, Anita Williams Woolley, and their co-authors at Carnegie Mellon, MIT, Illinois, and Harvard put the alternative in PNAS Nexus this spring. Organizations frame the issue as humans versus AI, they argue, but the better question is how to design teams “so AI expands what people can notice, remember, and reason through” while people provide “context, judgment, and accountability.”

What to say to your CEO this week. “Before we sign the next platform deal, I want a meeting on what we’re conceding by signing it. Not features. Not price. The decisions about how our people work together that are already inside the product. We’re three years in and we’re still acting like the question is which agent. The question is which assumptions we just locked ourselves into.” Then bring the assumption sheet (next section). It’s a thirty-minute exercise. Do it before the next platform pitch, not after.

Subscribe now

This week’s move: write the assumption sheet before the vendor sheet

The next agent-platform pitch will come with the standard procurement diligence. Data residency. SSO. Audit logging. Pricing. SLAs. The procurement function does fine work on those. It can’t do the work upstream of them, because the assumptions are baked into the product before any clause in the contract gets negotiated.

So write the assumption sheet first. One page. Six rows. One column for the question.

Coordination. Does this platform treat the coordination between our people as overhead to be routed around, or as work worth preserving? Look for shared visibility, joint review, paired sessions. If they’re not in the product, the product has a position on whether that work matters.

Unit of work. Does the platform decompose work into discrete tasks with clean handoffs, or does it support two people puzzling over an ambiguous case for an hour? Make the vendor demo the second one. Watch carefully.

Destination. When the vendor describes the end state, is a human still in the workflow or escalated out of it? Note the language exactly. “Truly autonomous” is a different bet than “agents as part of the team.”

Relational persistence. Does the system remember a person, or only a workflow? Memory Bank remembers expense habits. Ask whether anything in the product remembers what your team is working on together, what’s changed since last quarter, who carries which expertise. The answer is almost always no. That’s a choice.

Naming. Is the AI described as a teammate, an assistant, a tool, a service? Each is a different theory of what humans are for. Pick the one your team can actually use.

Exit. If this turns out to embed assumptions you’d want to reverse later, what does reversing cost? Get the answer in writing.

Bring the sheet to the next platform review. The questions are obvious once you see them, and almost nobody is asking them. That’s the chief of staff’s specific contribution.

Last time around: the divergence of 1985

October 1985, Cambridge, Massachusetts. The MIT International Motor Vehicle Program begins a five-year study across fifteen countries, led by James Womack, Daniel Jones, and Daniel Roos. The provocation is simple: Japanese auto plants are producing vehicles with half the defects and half the labor hours of US plants, using the same machines and many of the same components.

The answer, when The Machine That Changed the World publishes in 1990: Toyota and its peers had built a different theory of work into the same equipment. Andon cords let any worker stop the line. Quality was built in upstream, not inspected at the end. The worker was seen as a source of improvement, not a source of variance. Same machines, different assumptions, different cars.

US manufacturers had access to the same equipment. They chose differently, year after year, until the gap had compounded for a decade and the choice became fixed.

Manufacturing is not cognitive work, and Japanese employment in 1985 was a different beast. But the lesson translates: when a new technology arrives, the assumptions in how it gets deployed are at least as consequential as the technology itself.

From the frontier

In Longmont, Colorado, a 367 square foot greenhouse at 5,090 feet of elevation has spent the last few weeks running an AI planner called Iris alongside a $5 ESP32 microcontroller. Iris writes bounded climate tactics: setpoint biases, mist limits, venting posture. The ESP32 owns the relays on a five-second cycle, with the safety logic local on the chip. This is the assumption sheet built into hardware. The owner publishes everything: plans, scorecards, costs, stress hours, water use, failures, and the exact list of tunables the AI is allowed to change. You can go and look at the live Grafana panels right now. Source

Potpourri

From someone doing it. Amy Hall, dean of nursing at Franciscan Missionaries of Our Lady University, told Axios this week that “nurses need a stronger voice in which tools are adopted.” She was responding to the Elsevier Clinician of the Future 2026: Nurses Edition report, released around International Nurses Day on May 12, surveying 692 nurses and 2,065 doctors across 118 countries. The headline: 41% of nurses said their views are rarely or never adequately represented in their organization’s decision-making, against 19% of doctors saying the same. Source

From the weirder edge of the week. At the SFER IK museum in Tulum, an installation called The Bat Cloud invites you to ask a bat a question. The wooded area around the museum has been seeded with water, fruits, and insects, and an AI oracle trained on bat vocalizations answers your question with the help of the bats’ actual responses to the environment. Source

04: We Have Some Concerns

Bud Caddell — Fri, 08 May 2026 19:31:54 GMT

This week’s main thing

This is the week the AI narrative hit the fan. Stanford’s 2026 Index reported the widest expert-public divergence it has measured. Gen Z anger about AI overtook their excitement in a Gallup survey of 1,572 14-to-29-year-olds. Anti-AI sentiment intensified in headlines from the Hollywood Reporter, Fortune, MIT Technology Review, and TechCrunch. Pentagon clearances, Wall Street rollouts, a unionization vote at DeepMind, an Anthropic compute deal with the Musk-built Colossus, and a literal kill-switch demo at the largest enterprise software conference of the year all landed in the same seven days.

A gloomy economy + AI washing layoffs + an entire industry that positioned AI as adversarial to humans to such an extent that one company brazenly runs “don’t hire humans” billboards = people are pissed and suspicious.

But here’s the thing: this perception gap has been brewing for years, and researchers have already explored how to respond.

If you inform how your org integrates AI, this is the week to tell your CEO that the perception of these tools matter and a listening tour isn’t going to cut it. Trust in leaders is at an all-time low, and a promise to listen (in words alone) will carry almost no weight with workers.

What you need is an AI Deployment Council.

What to say to your CEO this week. "I want to propose we establish an AI Deployment Council this quarter. Here (below) is a charter we could start from. The conversation worth having is what scope and authority we are willing to give it, not whether to have one. The current trust environment will not let an announcement or vague promise do the work and we could be heading off an outright revolt."

This week’s move: propose an AI Deployment Council

Propose your company form an AI Deployment Council this quarter. Not a town hall, not a pulse survey, not an executive-led “AI ethics committee.” A standing body with workforce representation, published scope, a regular cadence, written response obligations from leadership, and public outputs. We have published a starter charter drawn from fifty years of European works council practice and the recent research on AI deployment governance. It is concrete enough to be lifted into a real proposal this week.

Expect four objections from your CEO.

“We already have employee feedback channels.” Pulse surveys are not consultation. The PwC data is the answer: 86% of executives believe employees highly trust them; only 67% do. Leaders systematically overestimate how their existing channels work. A council with written response obligations and published outputs is structurally different.

“This will slow us down.” Yes, in the same way reversal triggers slow you down, which is the point. Speed without consultation is what produced the DeepMind union vote and the Project Maven retraction. Speed with consultation is what produced fifty years of higher productivity in German co-determined firms.

“This sounds like a union.” It is not a union. It is a consultative body that exists alongside any union representation. The deeper answer: workers organize unions when they do not have a council. DeepMind is the warning. The choice is between offering structure now or having a more adversarial structure forced later.

“We cannot give workers a veto.” This is the real conversation, and it deserves an honest answer. The charter offers a spectrum: pure advisory, consultative with written response obligation, or co-determination on a few narrow categories. Most US companies in 2026 will start at consultative-with-response-obligation, not co-determination. That is meaningfully different from theater (the response is public and binding) without giving anyone a veto.

What the council cannot do: set strategy, make financial decisions, or override customer-facing product choices outside workforce impact. Be honest about scope. The credibility of the council depends on what it actually does, not what it is announced to do.

Last time around

April 11, 1812. A Saturday, just past midnight. More than a hundred men gathered silently at the Dumb Steeple at Cooper Bridge, a stone obelisk by the road in West Yorkshire. They marched in military formation to Rawfolds Mill, organized by company, called by numbers rather than names. The mill belonged to William Cartwright, who had installed mechanical shears that did the work of skilled croppers, men whose trade required a seven-year apprenticeship. Cartwright had been warned about the attack. He had a detachment of soldiers garrisoned inside, beds for his men set up in the counting house, and loopholes cut in the walls.

The garrison opened fire. The fight lasted twenty minutes. Two attackers were mortally wounded; one was John Booth, a 19-year-old apprentice cropper, tortured for names by a magistrate who refused him medical attention until he talked. Booth died without giving any.

The British state deployed 12,000 troops to the Midlands to counter The Luddites, more than Wellington had in Spain, and made machine-breaking a capital crime. Seventeen Luddites were hanged at York Castle the following January. Skilled cropper wages collapsed. The Factory Acts that addressed the worst abuses came forty years later, after the displaced workers were dead.

E.P. Thompson called what happened at Rawfolds “collective bargaining by riot.” The Luddites were not anti-technology and not irrational. They were skilled workers whose every legitimate channel, Parliament, the courts, the guild structures, the master-apprentice relationships, had been closed or rendered inert. Machine-breaking was what was left.

The lesson for 2026 is not that revolts are coming. It is that when the formal channels for affected workers to shape deployment close, the informal ones expand. The DeepMind union vote, twenty data center projects blocked or delayed, lawsuits over training data – these are all attempts to use formal channels to influence the direction of a powerful new piece of technology. Spen Valley Civic Society

From the frontier

Four physicists at Emory, Wentao Yu, Eslam Abdelaleem, Ilya Nemenman, and Justin Burton, used a custom neural network to derive new physical laws governing dusty plasma, the ionized gas filled with charged particles found in everything from Saturn’s rings to wildfire smoke. The humans designed a network with built-in symmetries, ran the lab experiments, and framed what would count as new physics. The AI did the pattern-matching across particle trajectories that traditional analysis could not crack, learning non-reciprocal forces between particles to better than 99% accuracy from a small dataset. The result corrected longstanding theoretical assumptions and produced a generalizable framework that may apply to other many-body systems. The paper ran in PNAS; ScienceDaily re-covered it in late April. PNAS

Potpourri

From someone doing it. Lori Beer, global CIO of JPMorgan Chase, told Fortune at Anthropic’s May 5 financial services briefing that the bank’s central problem with AI is not the technology itself. “There’s this capability overhang. The technology can do so much. It’s the actual organization’s ability to digest and absorb it that tends to be where the gap is.” When the global CIO of a 319,000-employee bank names organizational absorption as the binding constraint on AI deployment, the conversation has moved past capability and into governance. Fortune

03: The Undo Button

Bud Caddell — Fri, 01 May 2026 19:58:24 GMT

This week’s main thing

The question this week isn’t “What can AI do?” It’s “What can humans undo?”

Most of the systems we’re deploying agents into are too tightly wired together for “undo” to mean what we think it means.

Quick rewind: In his 1997 letter to shareholders, Jeff Bezos drew a distinction Amazon still uses: some decisions are two-way doors and some are one-way doors. A two-way door, you walk through, look around, walk back out. A one-way door closes behind you. Bezos argued that organizations should make two-way-door decisions fast and one-way-door decisions slowly, with deliberation.

Our AI needs an analogous ruleset. But it’s complicated.

Sociologist Charles Perrow studied disasters at Three Mile Island, chemical plants, and in aviation, and found that the type of door (one-way or two-) is a property of the system, not the decision itself. Some systems are “tightly coupled” – where when one action is taken, its new state triggers additional actions and downstream consequences at a rate too quick to reasonably stop.

Imagine a platform agent that closes a marketplace seller’s account due to fraud signals: a two-way door from the admin panel, a one-way door once the closure has cascaded to active orders, refunds, search ranking, and external syncs that ran in the same few minutes.

So the question for AI integration this week isn’t “do we have a rollback path?” It’s the harder question Perrow forces: how tightly coupled is the system this agent is acting in, and at what point does that coupling turn the door one-way regardless of what the policy says?

Here’s a start, drawn from how Perrow’s framework gets applied in high-reliability fields like aviation and hospital medication systems:

Low coupling. The agent’s action affects one record. Other systems read it on their own schedule. Reversal is straightforward. Agents act, with logging.
Medium coupling. The action triggers downstream behavior within minutes, not seconds. Reversal is possible but requires compensating actions. Agents recommend and even act under a subset of conditions; but humans have to approve in the majority of use cases.
High coupling. The action triggers automated downstream behavior in real time, where customers, regulators, or partners see the result before we do. Reversal is theatrical. Agents surface the option; a human walks through the door every time.

What to say to your CEO this week

Before we sign off on the next agent deployment, I want to flag something. The way we’re scoping these is by what the agent can do, but the question I’m not seeing us ask is what’s around the action when it happens. There’s a useful frame from Bezos here: two-way doors versus one-way doors. We’re set up to identify those when humans are walking through them, because humans pause and read the sign. Agents don’t. They walk.

I want us to sort our active and planned agent deployments into three buckets. Where the agent’s action affects one record and the world catches up later, agents can act. Where the action triggers downstream systems within minutes but not seconds, agents recommend and even act in specific circumstances, and a person commits the rest. And where the action triggers automated downstream behavior in real time, where customers, regulators, or partners see the result before we do, the agent should surface the decision but not make it. Not because we don’t trust the AI. Because there’s no door to walk back through. And the EU’s Article 14 enforcement on August 2 is going to ask a version of this same question with regulatory teeth.

Subscribe now

This week’s move: sort your agent actions by coupling, not by capability

Most agent deployment reviews ask the wrong question first. They ask what the agent can do. The better first question is what’s around each action when it runs. Once you have that, the right level of human involvement falls out of it.

Run this on the most-deployed agent in your organization. List every action it takes. For each one, ask three questions in order.

One: what else happens when this action takes place? If nothing else moves for at least an hour (the action writes to a record other systems pick up on a daily sync), the coupling is low. If three other systems respond within minutes (notifications, calculations, status updates), it’s medium. If real-time downstream actors respond (other agents, customer-facing state, partner APIs, regulatory feeds), it’s high.

Two: who sees the result before we do? If the only audience is internal staff who can flag a problem, low. If it includes customers who’ll see it in their next interaction with us, medium. If it includes the customer’s customers, regulators, credit bureaus, or anyone outside our company whose response is automated, high. The faster the external audience reacts, the more one-way the door.

Three: is our time-to-detection shorter than the cascade? If we’d notice within an hour and the cascade takes a day, the coupling is loose enough to recover. If we’d notice within a day and the cascade completes in an hour, the door has already closed. The Flash Crash, in the analog below, did $1 trillion in damage in 23 minutes. No human noticed in time because the system’s clock was faster than the humans monitoring it.

Last time around

May 6, 2010. 2:32 PM Eastern. A mutual fund called Waddell & Reed sells a $4.1 billion block of stock-market futures through an automated trading program. The program is told one thing: feed the order into the market based on how much trading is happening, with no price floor and no time limit. Volume rises. The program responds by selling more. Other automated programs running their own logic interpret the selling as a signal and start selling themselves. By 2:45 PM, the Dow has fallen nearly 1,000 points. The market loses roughly $1 trillion in value in less than half an hour. Accenture briefly trades at one cent. Recovery is partial by 3 PM.

The fix is not a smarter algorithm. It is single-stock circuit breakers, approved by the SEC in June 2010 and rolled out to all 404 S&P 500 stocks by mid-June. The first one trips on June 16, on Washington Post Company shares, after three erroneous trades cause a price spike. The breakers do not predict bad behavior. They sit one layer below the algorithms, in the architecture of the market itself, and they make certain failure modes mechanically impossible.

Markets had been algorithmically traded for years before May 6. The infrastructure that lets the market recover from an algorithm gone wrong arrived after the trillion-dollar incident, not before it. The agentic AI deployments shipping this quarter are running in front of their circuit breakers. SEC/CFTC Joint Report

Potpourri

Birgitta Böckeler outlines a proper harness. “Harness engineering” is a term that emerged in February for the layer of guardrails, checks, and review steps built around an AI agent. Böckeler’s essay is a great primer on what’s necessary to guide and constrain agent behavior (e.g. feedback and “feedforward” controls). The questions that emerge: what’s the role of a human in oversight, how much can one agent regulate another, and how much control will we ever have over these non-deterministic creatures. Martin Fowler

02: Weird Flex

Bud Caddell — Fri, 24 Apr 2026 17:29:49 GMT

This week’s main thing

This week, AI company CEOs are telling reporters that half of entry-level white-collar jobs will be gone in one to five years, and tech companies are laying people off in large numbers blaming AI. At the same time, the researchers studying what AI is doing inside organizations, including those employed by those same AI companies, keep publishing careful findings that AI is reshaping roles, not deleting them.

Putting aside for a minute that the clearest ROI so far for AI is its ability to give cover for a layoff, it’s worth asking this week just how flexible jobs really are (i.e., how much can a stack of tasks or roles really bend before a human being is broken and/or a job is deleted).

Here’s what we know from org dev research:

There are 3 dimensions of job flex and AI hits them unevenly: tasks, relationships, and meaning.

Tasks: this one is the easiest to address, and you can think about the set of tasks someone has as either tight or loosely bundled based on how interrelated or interdependent they are with each other. Some people (often junior staff) have a loose pile of things they do that really don’t have a ton to do with one another, they are just things that need to get done. AI is actively destroying roles with loose bundles. Other folks, often more senior, have a bunch of tasks all related to the same core thing that are dispersed along different cognitive abilities (some are repetitive, some require deep thinking). Because AI isn’t galaxy-brain yet and struggles to hold context for long periods of time, these roles can adapt as AI eats their repetitive tasks.

Relationships: this is who you need to work with to get your tasks done and AI is making this one really hard to predict because it’s hitting all roles in the org chart simultaneously AND it’s making work feel more and more like a single player game. Turns out though, we bond at work through shared problem solving and so we need AI to leave us alone to struggle together at some points. Also, junior folks need the struggle to build both their social networks and their competence.

Meaning: as a society, we are running at job replacement without thinking this one through–if all we’ve left for people to do is click “approve” or “next” then this role has no meaning, and you can basically assume people will do it poorly. Trying to preserve the role then becomes self-defeating.

So when you’re evaluating a person’s role, consider how loose or tight their task bundle is, which relationships they really should be retaining and why, and if you’ve left them any point to getting out bed in the morning.

But here’s the kicker: flexing itself is not what breaks people. Rate is. The change-fatigue research says humans can absorb substantial role redesign if they get six to eight weeks of stability between moves. They give up when they’re given no time to settle into a new role.

What to say to your CEO this week:

Before we sign off on the next round of AI role redesign, I want to flag something. Our people are already absorbing more change than the research says humans can handle without a break. Gartner has the average employee at 14 concurrent changes, up from five in 2016, and willingness to go along with change has collapsed over the same period. Six to eight weeks of stability between moves is roughly the threshold where adaptation keeps working. I don’t think we’ve given anyone that this year.

The AI redesign itself is the right direction. But if we layer it on top of everything else that’s in flight, we aren’t going to get the productivity gain we’re modeling. We’re going to get attrition from the people we most need to keep, and quiet disengagement from the rest. I’d like us to look at the change calendar before we set the timeline on this one.

This week’s move: scenario work beats forecasting when the horizon is short

A common complaint this month from people trying to write anything serious about AI is that the three-year horizon has collapsed to eighteen months. Strategists keep running forecasts that age out during the quarter they were built for. Technology assumptions that ground the analysis change inside the analysis window. Executives feel this too. The usual response is either to ignore the uncertainty and forecast anyway, or to freeze. Both are bad.

The move is structured scenario work, and speculative fiction is better raw material for it than most strategy consultants will admit. Shell’s scenario planning team, under Pierre Wack in the early 1970s, used fiction-adjacent thinking to prepare the company for the 1973 oil shock while competitors were running straight-line forecasts. The point was not to predict the shock. It was to have already imagined a world in which it happened, so that when it did, the organization had mental infrastructure to respond.

A useful new resource for this, surfaced this month, is the Extrapolated Futures Archive (urubos.github.io/efa-site). It catalogs 276 science fiction ideas mapped to 1,903 stories, tagged by domain, scenario type, and outcome. A reader can search by situation and get a ranked list of fictional precedents. It is built, its creator says, for decision-makers who want to widen their thinking before a decision rather than after.

A chief of staff preparing for a leadership offsite this quarter might try a one-hour version of this. Take three of the AI decisions your company is currently treating as settled. For each, pull two or three fictional treatments of analogous situations (creation escaping creator control, automation and labor displacement, all real EFA entries). Have the leadership team argue the decision as if the fictional situation were the actual world. Not to predict. To pressure-test what the decision assumes. The output is a list of assumptions you can now watch the data for.

Last time around

February 4, 1985, 12:05 PM. The first car rolled off the line at GM’s Detroit-Hamtramck plant, a Cadillac Eldorado, delivered with fanfare. Roger Smith had spent $500 million on the factory and bet GM’s future on what a company brochure called “the passage to the future”: 260 robots, 2,000 programmable devices, an automated dolly system that moved parts around the floor without human hands. Part of a larger “Factory of the Future” program meant to grow GM’s robot count from 302 units in 1980 to 14,000 by the end of the decade.

The dolly wandered off course. The spray-painting robots sprayed each other. For months, Hamtramck trucked half-finished cars across town to a 57-year-old Cadillac plant so human workers could repaint them. When the robogate welding machine smashed a body, or a welder stopped dead, the whole line stopped, and workers stood around while managers called the robot vendor’s technicians. Paul Ingrassia and Joseph White, reporting on it later, wrote: “They had simply grafted robots onto the old, inefficient system. GM bet the entire Hamtramck production system on the proposition that leading-edge automation would work instantaneously.”

It didn’t. F. Alan Smith, GM’s former CFO, quipped in 1986 that the company would have been better off using the money to buy its two biggest competitors outright. GM’s market share fell from 46% to 35% over the decade. Toyota spent the same period putting fewer robots next to more capable workers and kept winning. Today, automobile assembly is largely automated. Smith had the right long-run direction, but the sequence cost him the company. Fortune

Potpourri

Dan Hon, writing in his newsletter Things That Caught My Attention on April 16, on why people say the economy feels worse than the data says: “I legit think some of the belief and experience that the economy sucks is also sheer exhaustion from shit like this and notification fatigue and every single shitty tech-mediated interaction of which there are hundreds a day.” His frame is convenience-for-agency. Every tech-mediated convenience is a small transfer of agency from the user to the platform, and they add up. Things That Caught My Attention

01: Metrics and Meat Shields

Bud Caddell — Fri, 17 Apr 2026 20:28:53 GMT

This is Superadditive’s first weekly newsletter, so let’s start with what it is:

These briefings are for the people inside organizations shaping how AI actually gets used.
We want this to be worth reading on your weekend—something you can sit with and show up Monday ready to act on.
Each week, we look for the signal inside the noise: where AI is making teams smarter, where it’s quietly degrading judgment, and where we’ve seen this pattern before.
There’s no shortage of newsletters. We aim to be the best read in your inbox. If we’re not, tell us. If we are, spread the word.

Takeaways from this week

A synthesis of our top stories and what it means right now

Three things happened in the same six days, and they are worth considering together.

On April 10, Duolingo CEO Luis von Ahn said on the Silicon Valley Girl podcast that the company would no longer factor AI usage into employee performance reviews. A year earlier, in April 2025, von Ahn had declared Duolingo “AI-first” and announced that employees would be evaluated, in part, on how much they used AI in their work. That last piece is what he is walking back. Employees had started asking whether they were being graded on their work or on their tool use, and he could not give them a good answer.

We also got two separate studies that demonstrate how complicated AI integration really is. First, researchers from UCLA, MIT, Carnegie Mellon, and Oxford released a preprint showing that participants given AI help on reasoning tasks performed worse, and gave up faster, once the AI was taken away. Second, the American Psychological Association published a peer-reviewed study of 1,923 workers showed that how people interact with AI outputs (whether they just accept it or edit it critically) affects how they rate their own ability to think.

Clearly, understanding how people are using AI is more important than just counting if they use it. But Shopify, Meta, Nvidia, and McKinsey are still pushing in the direction Duolingo just came back from. Meta requires engineers to hit a quota of agent-assisted code changes. Jensen Huang told Nvidia engineers earning over $500,000 that he expects them to burn through at least $250,000 in AI tokens per year. These are the modern equivalent of lines-of-code quotas. They will produce exactly what such quotas always produce, which is more of the thing being measured and less of the thing the measurement was supposed to be a proxy for.

What to say to your CEO this week: Our peer companies are installing AI-usage targets. The newest evidence says the targets measure the wrong thing. The research that came out this week shows that outcomes depend on whether people edit and push back on what AI produces, which is the opposite of what a usage quota rewards. If we need an AI metric for the board, make it about work quality with AI help, not tool use. Duolingo, a year ahead of us on this, just reversed course.

This week’s move

If your board or CEO is asking for an AI metric, the question to run at them is: would we have accepted this metric for a software tool in 2019? Nobody measured Excel usage. The tool is a means. The work is the end.

A usable AI metric has three properties. It measures outcomes, not inputs. It is resistant to gaming. And it can be computed from work that already exists, not from a new surveillance layer.

Three candidates that clear this bar. Rework rate — the share of first-draft AI-assisted output that has to be substantially redone downstream. The Stanford and BetterUp “workslop” research puts this at around 15% of all work content in the average organization, with receivers spending close to two hours per instance cleaning it up. If your rework rate is above that, your AI is not yet a net positive. Cycle-time on well-defined tasks — the time from request to shippable deliverable on tasks with clear specs, compared to the same cohort of work last year. This one is gameable but only in directions you want (smaller tasks, cleaner specs). Error-catch rate — in regulated or safety-critical workflows, the share of AI-suggested outputs that a human reviewer changes before release. A high catch rate is good. A catch rate of zero means your reviewers have stopped looking.

None of these require new tooling. All three survive the question “what would this incentivize?”

Avoid: token spend, prompts per day, hours logged with an AI tool, percent of code or copy AI-generated. These are body counts. They look precise. They reward exactly the wrong behavior, which is the finding from the Dubey study this week.

“History does not repeat itself, but it does rhyme.”

In 1961, Robert McNamara arrived at the Pentagon from Ford, where he had been part of the Whiz Kids group that used statistical process control to turn manufacturing around. He brought the same instinct to the Vietnam conflict. The problem was that the war did not have a tonnage or defect-rate to optimize. So McNamara’s team picked a metric that could be counted, which was enemy dead, and made it the primary measure of progress. Body counts were published daily.

By 1977, Army historian Douglas Kinnard surveyed the generals who had commanded in Vietnam. Only 2% of them considered the body count a valid measure of progress. One called it “a fake, totally worthless.” Another wrote that the numbers were “grossly exaggerated by many units primarily because of the incredible interest shown by people like McNamara.” The Army had won on the metric and lost on the war.

Sociologist Daniel Yankelovich described the McNamara fallacy in four steps. Measure what is easy to measure. Assign arbitrary values to what is not. Assume what is not measured is unimportant. Conclude that what is not measured does not exist. Yankelovich called the final step suicide.

The parallel to the current moment is close enough to be uncomfortable. The CEOs installing AI-usage quotas are usually doing it because a board is asking how the AI investment is going. The measurable answer is time spent, tokens consumed, and percent of output AI-assisted. The unmeasurable answer — whether the work got better, whether judgment was preserved, whether institutional knowledge was retained, whether customers could tell the difference — is the one that matters. McNamara’s original sin was not bad math. It was confusing what was countable with what was important. Duolingo, a year into running its own body count, seems to have figured this out. The question is whether the companies that follow them will need a full cycle to learn the same thing. MIT Technology Review on McNamara and Kinnard’s survey, 2013; Yankelovich’s four-step framing via Wikipedia

Potpourri

From the floor. The Stanford and BetterUp research group that coined “workslop” — AI-generated content that looks polished but does not actually advance the work — ran a follow-up round surveyed this spring. In their original study, 40% of 1,150 full-time employees reported having received workslop in the past month, and receivers were spending close to two hours of cleanup per incident. A finding that has stuck with practitioners: roughly half of recipients said they now considered their co-workers less creative and less reliable after receiving AI-generated work from them. The productivity drain is measurable. The trust drain is worse. Harvard Business Review, September 22, 2025

Overheard. A Forrester researcher, quoted in the firm’s 2026 Future of Work report, on why so many companies regret AI-driven layoffs: too many executives “lay workers off for the future promise of AI.” The problem is not that AI cannot do the work. The problem is that it often cannot do it yet, and the cost of waiting is being paid by the people who were let go. Computerworld, November 4, 2025

Practitioner voices

Kyle Kingsbury, a distributed-systems engineer who runs the Jepsen consultancy and writes at aphyr.com, published a long essay on April 15 titled “The Future of Everything is Lies, I Guess: New Jobs.” One section anticipates a role Kingsbury calls the meat shield: the human whose job is to be accountable when an automated system fails. He writes that when Meta employs human reviewers to check automated moderation decisions, or when lawyers get sanctioned for submitting AI confabulations to a court, the company is “dangling a warm body over the maws of the legal system and public opinion.” You can fine a corporation that uses an LLM, but only a human can apologize or go to jail.

Kingsbury connects this directly to Madeline Clare Elish’s concept of the moral crumple zone, where the human operator absorbs legal and moral blame for a mostly-automated system’s failures. The framing is useful because it names something that is showing up in the quiet rewrites of employment contracts and job descriptions: accountability is being pushed down even as decision authority is being pushed up to the model. aphyr.com, April 15, 2026

Fieldnotes from Superadditive

07: Who is this for?

This week’s main thing

This week’s move: make decisions as if you had to do explain it face to face

Top stories

Last time around

Potpourri

06: The Blind Firing Squad

This week’s main thing

This week’s move: How to Conduct AI-Related Layoffs

Top stories

Last time around

Potpourri

05: Japanophilia and Greenhouses

This week’s main thing

This week’s move: write the assumption sheet before the vendor sheet

Top stories

Last time around: the divergence of 1985

From the frontier

Potpourri

04: We Have Some Concerns

This week’s main thing

This week’s move: propose an AI Deployment Council

Top stories

Last time around

From the frontier

Potpourri

03: The Undo Button

This week’s main thing

What to say to your CEO this week

This week’s move: sort your agent actions by coupling, not by capability

Top stories

Last time around

Potpourri

02: Weird Flex

This week’s main thing

What to say to your CEO this week:

This week’s move: scenario work beats forecasting when the horizon is short

Top stories

Last time around

Potpourri

01: Metrics and Meat Shields

Takeaways from this week

This week’s move

Top stories

“History does not repeat itself, but it does rhyme.”

Potpourri

Practitioner voices