Conductors and Casualties
The agentic moment is reshuffling labor along a new axis. The bet is partly right. The casualties are not abstractions.
Last week, Coinbase reduced its workforce by roughly 14%. In the email Brian Armstrong sent his employees, he did not soften the rationale. He stated it plainly. The company, he wrote, would rebuild itself as “an intelligence, with humans around the edge aligning it”.
That sentence deserves to sit for a moment.
It is the clearest articulation publicly available of what businesses are now betting. Capital is moving from human labor to AI on the assumption that productivity will rise enough to absorb the displacement. The intelligence is the substrate; the humans are the alignment layer at the perimeter. That is a topology. It is a claim about where the center of gravity in the work now lives.
Coinbase is one company, on one date, with one CEO willing to say the quiet part on the record. It will not be the last. The pattern is the story, not the company. Within the next twelve to eighteen months, dozens of mid-cap and large-cap companies will publish similarly worded restructuring announcements. Some will use Armstrong’s language directly. Most will use a softer translation. The bet underneath will be the same.
The bet is partly right and partly wrong. That is the thing worth getting clear about, because most of the current commentary refuses to hold both halves at once. The optimists insist the bet is fully right. The skeptics insist it is fully wrong. Neither posture is honest about what the tools can do and what they cannot. Neither posture is useful to a worker in week three of trying to figure out whether the next round of layoffs has their name on it.
The part the bet gets right is more uncomfortable than the skeptic-case allows. The part it gets wrong is more structural than the optimist-case admits. And the gap between the opportunity the moment opens and the formation most workers have received to meet it is the place where actual people are going to get hurt.
What the bet gets right
The productivity gain is real. Not hype. Not a narrative inflation by people with equity in the outcome. Real.
Here is one data point from this past week. A long source document, an aesthetic exemplar, a single key color choice, plus a request for fifteen feature images tailor-made to the content. My hands-on time across the whole project came to fifteen to twenty minutes. The rest was wait time on the image generator, which ran in the background while I did other work. Commissioning those images traditionally would have taken days, possibly more than a week, and cost hundreds of dollars on the low end. That kind of compression is happening across knowledge work, in real teams, in companies you have heard of, this week.
The external existence proof is sharper. Andrew Wilkinson is running a SaaS at $20,000 in monthly recurring revenue with zero employees. Support, marketing, development, all handled by Claude Code agents. One human at the center, several agents at the edges. Greg Isenberg surfaced the structure for the people who had not yet seen what was coming. The structure works. It is not a thought experiment. There is a real customer base paying real money to a company with one human in it.
The skeptic case that “this is just hype” is not the right argument. The right argument is more uncomfortable: the productivity gain is real, and that is exactly why the labor-market reshaping is going to happen regardless of whether anyone thinks it should. Markets do not consult our preferences. When a tool genuinely compresses a four-hour task into a forty-five-minute one, capital flows toward the compression. That is not a moral claim. It is a description of how capital behaves in the presence of a productivity differential.
So grant the bet its full weight. The compression is real. The structure works. The reallocation is rational on the metric the reallocators are using.
The question is whether the metric is the whole picture.
What Apple’s paper actually shows
In June 2025, six researchers at Apple’s Machine Learning Research division published a paper titled The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity. It is the most important empirical document of the current AI window for anyone trying to think clearly about where the human belongs in the system.
The paper maps three distinct performance regimes, and the nuance matters. On low-complexity tasks, standard LLMs outperform the newer reasoning models. On medium-complexity tasks, reasoning models gain advantage. On high-complexity tasks, both classes collapse completely. The authors describe “complete accuracy collapse beyond certain complexities”. Not degradation. Collapse.
The most counterintuitive finding, and the most important one, is about effort. The authors observe that “reasoning effort increases with problem complexity up to a point, then declines despite having remaining token budget”. The models use fewer tokens on harder problems. Not more. The opposite of what reasoning, in any sense the word usually carries, would predict. A human who hits a hard problem leans in. The model leans out. That is a description of imitation, not cognition.
The other finding worth quoting verbatim: the models “fail to use explicit algorithms”. Even when researchers gave the system the algorithm, spelled out, the system did not consistently apply it. That is a striking result. It is the kind of finding that has to be read slowly to feel its weight.
Methodology critiques exist, and intellectual honesty requires identifying them. Some researchers have argued the token-use decline partially reflects output truncation rather than reasoning collapse. Tower-of-Hanoi-style puzzles measure a narrow band of cognition. Humans also fail on scaled puzzles. None of those critiques are crazy. But the argument here does not require Apple to be unambiguously correct. Even granting partial pushback, the scale-failure pattern is sufficient to ground what comes next.
What comes next is the move almost no one writing on this paper is making. If the tools imitate rather than reason at the edge of complexity, the human reasoner does not become less load-bearing in the system. The human reasoner becomes more load-bearing. Not because the human is faster. Because the system needs someone whose cognition does not collapse when the problem gets hard.
That is a structural inversion of Armstrong’s topology. He put the intelligence at the core and humans at the edge. The paper suggests the inverse is closer to the truth.
The coordination tax
Before asking what AI can do for human productivity, it is worth identifying what is already consuming it. The coordination tax is documented, and the data are not soft.
Microsoft’s 2025 Work Trend Index, which is telemetry from real Microsoft 365 usage rather than survey self-report, found that knowledge workers are interrupted every two minutes during core work hours. Roughly 275 interruptions per day. 117 emails received. 153 Teams messages, up six percent year over year. Half of all meetings land in the 9-to-11am and 1-to-3pm peak focus windows, which is to say, in the windows where focus is most valuable, the calendar is most invaded. Sixty-eight percent of workers report not having enough uninterrupted focus time.
Asana’s Anatomy of Work Index puts the underlying split with painful clarity. Knowledge workers spend 60% of their day on what Asana calls “work about work”: chasing status updates, attending meetings about meetings, hunting documents across tools. Only 25% of the day goes to skilled output. Thirteen percent goes to strategic planning. Per worker, per year, the loss runs to 103 hours in unnecessary meetings, 209 hours on duplicative work, 352 hours talking about work rather than doing it.
The popular number, “the average worker is productive less than three hours of an eight-hour day,” comes from a Vouchercloud survey of UK office workers reporting two hours and fifty-three minutes of real productivity per day. The caveat is worth flagging: self-report surveys on productivity are notoriously unreliable. People underestimate and overestimate in different directions. The Vouchercloud number is a directional signal, not a precise measurement, and the methodology does not bear the same weight as Microsoft’s behavioral telemetry.
Different methodologies, same shape. The average knowledge worker is operating well below their productive ceiling, and the gap is consumed by coordination overhead, not by the work itself. That is the actual opportunity of the agentic moment. Not replacing the human. Freeing the human for the work only the human can do, by absorbing the coordination tax.
Which is a different vision of the work than the one Armstrong articulated.
The conductor opportunity
Genesis 1:28 is the dominion mandate. “Be fruitful and multiply; fill the earth and subdue it; have dominion over the fish of the sea, over the birds of the air, and over every living thing that moves on the earth” (NKJV). The mandate is not passive. It is active governance of a created order on behalf of a Creator who delegated the work.
Genesis 2:19-20 is the first recorded act of that governance. “Out of the ground the Lord God formed every beast of the field and every bird of the air, and brought them to Adam to see what he would call them. And whatever Adam called each living creature, that was its name. So Adam gave names to all cattle, to the birds of the air, and to every beast of the field” (NKJV).
John Lennox, in 2084 and the AI Revolution, treats this passage as more load-bearing than a casual reading suggests. “Naming things is a fundamental intellectual discipline in every area of inquiry,” he writes. “At its best, it involves not simply labeling them for convenience but understanding something of their nature.” For Lennox, the naming task was the way God taught Adam that humans are fundamentally distinct from animals. He adds, equally pointedly, that “humans are similarly fundamentally different from machines.”
The text frames this work as the first task. From the beginning, the image-bearer is the namer, the meaning-assigner, the one who reasons under delegated authority and exercises judgment over the system.
AI joins a long lineage of tools the image-bearer has made and directed. It does not displace the vocation. It tests whether the vocation is still being exercised.
Here is the metaphor the moment calls for. A conductor does not play every instrument. The conductor holds the score, hears the whole, directs the execution, and makes the judgment calls that no individual instrument can make from its position in the ensemble. The musicians execute with precision. The conductor provides direction, judgment, and critical thinking. The ensemble is greater than any of its parts because someone is holding the whole.
That is the topology the agentic moment calls for. Human at the reasoning core. AI at the execution edge. Coordination overhead absorbed by the system. Judgment, direction, evaluation, and articulation of intent held by the person.
It is the inverse of Armstrong’s phrasing. He put the intelligence at the core and the human at the edge. The conductor model puts the human at the core and the intelligence at the edge. Same labor, opposite topology. The Christian dominion diagnostic in Tools, Not Taskmasters identifies the structural commitment. The argument The One Thing No Algorithm Performs traced, that the human in the system is doing something the system cannot, is the cognitive anchor for what the conductor does that the agents cannot.
Not fewer humans. Humans differently positioned.
Four months in the rooms
For the past four months, regular leadership coaching sessions at work have focused on one question: how to use AI well. The participants are organizational leaders, often pastors of some of the largest churches in the country. The sessions are not theoretical. They involve hands-on work with the tools, in the leaders’ actual workflows, on the leaders’ actual problems.
I have been following the development of these tools since before ChatGPT released, and using them effectively for many hours a week since. What I have to report from these sessions is observed, not theorized.
The work in those meetings has two parts. The first is showing leaders what the current tools can actually do, in their workflows, on their problems, moving the conversation past abstractions into examples they can hold. The second is describing how this technology will reshape business, administration, and leadership in ways that are coming whether anyone in the room is ready or not.
At some point, a new question comes up. Not just how the leader should use these tools, but how to help the staff team learn to use them. The same hands-on work, extended to the people who carry the daily operating weight of the organization.
Across these engagements, two distinct groups have emerged.
The first group is getting it. They have made the posture shift. They come to the session with outputs their agents produced. They bring questions about how to direct the next iteration. They have a growing instinct for what the system can do well and what it cannot do at all. They are becoming conductors. The work is visibly different. The output volume is higher. The quality of their judgment calls is sharper because they are spending less time on coordination overhead and more time on the reasoning work the moment requires of them. Their meetings are shorter. Their decisions are clearer. Their teams are noticing.
The second group is stalling. This is the harder observation to make, and it deserves precision.
The stalling is not about intelligence. These are sharp people. It is not about effort. They are showing up. It is about a specific kind of attention the work requires, and that kind of attention has, for some, atrophied. The work asks the leader to evaluate output, iterate on direction, and hold the whole in mind while directing the parts. It asks the leader to be the conductor rather than the first-chair player. For leaders whose entire competence has been built around personally executing the work to a high standard, the conductor posture is unfamiliar and uncomfortable. The instinct to do the work yourself fights the discipline of directing the work AI agents.
The pattern in The Expertise Trap is the right diagnostic. The leaders most likely to stall are often the ones whose identity is most fused with their function. When the function can be delegated, the identity has nowhere to stand. The stalling is not laziness. It is disorientation. Real disorientation, the kind that takes weeks to work through, not minutes.
The observation is not a judgment. It is a pattern. Some are getting it. Some aren’t. And the pattern has implications, because it does not stay inside the coaching room.
The casualties of the workshift
Most workers will not make this transition quickly. The conductor opportunity is real. The window to reach it is not unlimited. And the casualties of the gap are not abstractions. They have names and mortgages and kids in school.
There are three structural reasons most workers will struggle to make the shift, and each one operates at a different level of the system.
The first is that the work is genuinely new. Prompt-evaluate-iterate-direct is not a skill any current job description trained for. The closest analog is management. But even experienced managers were trained to manage people, not systems. The feedback loops are different. People push back; agents do not. People bring tacit knowledge; agents bring statistical regularity. The instinct for when to trust the output and when to push back hard is not transferable from prior experience. It has to be built, slowly, through reps. And the people most under pressure to deliver right now are the people who have the least bandwidth to invest in learning a new posture.
The second is that reward structures still pay for doing-the-work-yourself. Most compensation systems, performance reviews, and promotion criteria are built around individual output. The person who directs an agent to produce ten times the output is not yet rewarded at ten times the rate. Often, they are mistrusted for not having done the work themselves. The person who does the work themselves, visibly, in the meeting, on the late-night Slack thread, is still the one who gets the credit. Until reward structures catch up to the new topology, the rational individual incentive is to keep doing the work yourself, even when delegation to an agent would produce a better outcome.
The third is that formation institutions have been building elsewhere. The skills the conductor moment requires—critical thinking, reasoning, articulating intent clearly, evaluating output against a standard, holding a whole in mind while directing the parts—are the skills education systems have been systematically de-emphasizing for a generation. The liberal arts were the training ground for exactly this kind of work. The market spent twenty years telling students they were not worth the tuition. We trained a generation away from the disciplines the moment now requires. The bill is arriving now, and it is large.
The casualties are not the people who refused to learn. They are the people who were never given the formation the moment requires, and who are now being asked to make the transition faster than any institution is equipped to help them make it. We did this. The schools we built, the reward structures we wrote, the cultural script we ran for two decades that said the only purpose of education was a higher-paying job. These were our choices. The bill is not arriving on someone else’s desk.
The question is not whether the workshift is happening. It is. The question is whether the people on the wrong side of it have anyone standing with them.
What the rooms are also for
The coaching sessions are not only about productivity. They are also about formation. The question underneath the question, how do I use AI well?, is always a question about what kind of person I am becoming in the process of using it. Some leaders have figured that out without anyone telling them. Some need someone to say it out loud.
Romans 12:2 (NKJV): “And do not be conformed to this world, but be transformed by the renewing of your mind, that you may prove what is that good and acceptable and perfect will of God.”
The renewing of the mind is not a metaphor for cognitive upgrade. Paul is not describing a productivity protocol. He is describing the formation work that produces a person capable of discernment: capable of knowing what to direct, what to question, and what to refuse. The conductor moment requires exactly that kind of person. Not a faster processor. A formed one.
The Church has always been in the formation business. It has always known that a person is not a sum of skills, that discernment cannot be downloaded, that the mind is renewed by sustained practice in a community of people committed to the same work. The market is now discovering, in its own register, that formation is the bottleneck. The tools work. The people are the variable.
Some of the largest churches in the country are starting to ask the same question, in their own register: what would it take to bring our staff teams through this work? The institutional motion is beginning. The Church does not need to chase the market to be relevant to this moment. It needs to do what it has always done, with clarity about why it matters now.
The market will not wait. The Church has always been willing to.
Sources
Brian Armstrong, Coinbase restructuring announcement (X, May 2026)
Greg Isenberg, on Andrew Wilkinson’s $20K MRR solo SaaS (X, May 2026)
Shojaee, Mirzadeh, Alizadeh, Horton, Bengio, Farajtabar, The Illusion of Thinking (Apple ML Research, 2025)
Shojaee et al., The Illusion of Thinking (arXiv 2506.06941, 2025)
Microsoft, Breaking Down the Infinite Workday: 2025 Work Trend Index
Vouchercloud, Survey Reveals Employee Productivity Averages 2 Hours and 53 Minutes a Day
John C. Lennox, 2084 and the AI Revolution, Updated and Expanded Edition (Zondervan Reflective, 2024)
Tools, Not Taskmasters: A Christian Posture Toward AI (April 2026)
This article was developed using AI writing tools I built to work with my voice, research, and editorial framework. The ideas, arguments, and theological positions are mine. The pipeline that helps me draft, evaluate, and refine them is something I created as part of my work at Nomion AI. I believe in building with AI and being honest about it. If you want to know more about that process, ask me.