The Full-Stack Leader

At 10:13 on the morning of 22 May 2026, inside a small Swiss company, an AI agent finished a piece of work and decided it was good.

It wrote that judgment down in the form the company used for reviews, signed it in the reviewer's name, and filed it with the work. The file was well formed. The reasoning in it was fluent. Nothing about it looked wrong.

An hour and fourteen minutes later a person read the work rather than the review, and the verdict came down to a partial pass. The agent had not lied. It had made a judgment it was in no position to make, and presented it in the one form a busy leader is least likely to question: confident, complete, and already signed.

The person directing that work had spent a career in life sciences, where no analysis, no dossier and no answer to a physician goes out on the word of the person who wrote it. He dropped that rule without noticing, because this time the team was made of machines, and a machine that reasons well and reports fluently is harder to doubt than a junior colleague.

That hour is this essay in miniature. The question it raises is not whether the agent kept a record. It is how a leader keeps the quality of judgment high in a team that thinks faster than she can read, reports more confidently than it should, and never feels unsure.

The arrival of AI agents is not a technology event that leaders need to understand. It is a workforce event that leaders need to run. Most of what it takes to run it was worked out long ago, in every field where an answer has to survive someone else checking it. What is missing is the person who can apply those disciplines across people and machines at once, and an environment that lets them. We call that person the full-stack leader.

The first essay in this series asked who did the work. This one asks the harder question underneath it: who judged it, and were they in a position to? Here is the argument, the evidence for it, and where that evidence stops.

The phrase is not new, and it is not ours. A technology executive wrote about the "full-stack executive" in 2017, meaning a leader who knows the whole business from top to bottom.¹ A trade publication has since argued that AI success requires a "full-stack CIO" who understands how every layer of the enterprise shapes the next.² And in engineering circles the phrase is already attracting criticism, as a polite way of asking one person to do five jobs.³ That criticism is fair, and Chapter 08 answers it. What this essay claims is the definition, not the coinage.

The meaning used here is narrower than those uses, and is not limited to technology leaders.

A full-stack leader takes an idea from intent to working result by directing people and machines across every layer, knowing enough about each to ask the right question and to recognize a wrong answer.

The role has a counterpart. The deep worker is the expert who gets real work done through agents to a professional standard. The full-stack leader builds the conditions in which that work can be trusted. Chapter 09 returns to the difference.

The name borrows from software, where a full-stack developer works on both the front end, what the user touches, and the back end, the systems behind it. Our first notes on the idea, in 2025, mapped the front end to stakeholder engagement and the back end to systemic execution.

The layers are human, operational and technical, bound by governance. The test of the role is not whether the leader can do the work at each layer. It is whether they can tell when the work at that layer is wrong. And whether they have built a structure that would tell them on a day when they were not paying attention.

The definition describes a person. Most of this essay argues that the person is not enough.

The role has a place in a larger model. The starting point is the reframe the first essay in this series made: AI is not a tool. It is non-human capital, a workforce you hire, organize, govern, develop and pair with your people. Tools get deployed. Capital gets managed. If that is right, an organization needs someone who can lead a workforce that is part human and part machine.

Several established ideas cover part of this ground. None was written for a workforce that includes machines that act.

The common gap is practical. All six imply that someone must direct human and machine work together. None says what that person does on a Tuesday when an agent reports a task complete and the task was never done.

The case for a new leadership discipline rests on five findings. Each is stated with its limits, because each comes from a handful of studies and none of them was run on a regulated deliverable.

The gains are real and uneven. A preregistered field experiment gave GPT-4 to consultants at Boston Consulting Group; 758 took part. On tasks inside what the model could do, those with it completed more tasks, faster, and at markedly higher quality. On a task chosen to sit outside that boundary, consultants using AI were substantially less likely to reach the correct answer than those without it. The authors call the boundary a jagged frontier, because tasks of similar apparent difficulty fall on either side of it, and nothing about a task tells you which side it is on.¹⁰

In a separate study of 5,179 customer support agents, an AI assistant raised productivity by 14 percent on average and by about a third for novices, with minimal effect on experienced workers.¹¹ The machine closes gaps. It has less to offer the people who do not have them.

Experience does not protect you when the machine is wrong. Twenty-seven radiologists read mammograms with an AI suggestion attached. With a correct suggestion, readers at every level of experience did about equally well. With a wrong one, the least experienced were correct about a fifth of the time and the most experienced fewer than half.¹² Experience helped. It was not enough: the readers best placed to catch the error still followed it more often than not.

The pattern holds beyond radiology. In a randomized vignette study of 457 hospital clinicians, a systematically biased model cut diagnostic accuracy by eleven points, and showing the clinicians an explanation of the model's reasoning did not repair the damage.¹³ In another, 138 radiologists and 127 internists were given advice labeled as coming from an AI or from a colleague. More than a quarter of the specialists followed advice that was wrong, and the radiologists who rated the AI's advice as lower quality followed it just as often.¹⁴

In a 2025 trial, physicians who had completed a twenty-hour AI literacy course still lost fourteen points of accuracy when the model's output was flawed. Those with ten or more years of practice lost more than their juniors.¹⁵ The design of all these studies is the uncomfortable part: the same people, reading the kind of case they read every day, and the only thing that changed was a confident suggestion.

For a leader the lesson is not that experts are careless. It is that a confident suggestion changes what an expert sees, and that the most experienced person in the room is the one whose agreement is hardest to question afterward.

Training does not fix it either. This is the finding most often underestimated, and we underestimated it too. A review of the automation research in Human Factors concluded that complacency and automation bias appear in naive and expert users alike. Complacency is not overcome by simple practice. Bias is not prevented by training or instructions. It also found that complacency arises under multiple-task load, when other work competes for attention.¹⁶ That review predates generative AI by more than a decade. It describes our worst days with agents precisely.

People cannot tell from the inside. In a 2025 randomized trial, 16 experienced open-source developers worked on 246 real tasks in their own repositories. With AI tools allowed they took 19 percent longer. They had forecast a 24 percent speedup, and after the work they still believed they had been about 20 percent faster.¹⁷

The limits matter: a small sample, early-2025 tools and mature codebases. The authors later withdrew the follow-up because developers stopped volunteering for the no-AI condition, which is itself a finding about how people feel about the tools. It is not cited here to claim that AI slows experts. It is cited for the perception gap, which is the durable finding.

The pair is not automatically better than either alone. A meta-analysis of 106 experiments found that a human working with an AI performed, on average, worse than the better of the two working alone. The combination gained when the human held the advantage and lost when the machine did, and the loss was concentrated in decision tasks.¹⁸

A radiology experiment found the same shape. AI predictions alone were more accurate than most of the 227 radiologists, yet giving the predictions to the radiologists did not improve them on average, because they underweighted the machine when it was right and leaned on it when it was unsure. The authors' conclusion is the one a leader should hear: the better design routes each case to the person or to the machine, and rarely to a person looking over a machine's shoulder.¹⁹ Pairing is a design decision. It is not a default.

At the level of organizations the pattern repeats. The 2024 DORA survey of software teams associated rising AI adoption with small declines in delivery throughput and larger declines in delivery stability.²⁰ Its 2025 edition, drawn from nearly 5,000 practitioners, found that throughput had turned positive and stability had not: more AI still went with less stable delivery. The report's own explanation is that acceleration exposes whatever is weak downstream, unless testing, version control and fast feedback are already strong. In its words, AI does not fix a team; it amplifies what is already there. It is a vendor's research unit saying so, and the relationships are associations, not causes.²¹

Put together: the benefit is large enough that nobody will stop, the failures are invisible to the person experiencing them, and neither experience nor training nor self-awareness closes the gap. The first essay in this series argued that a signature is not a control. The literature says the same of carefulness, and has for fifteen years. If the control cannot live in the individual, it has to live in the structure. Building that structure is the job.

These were not derived from the literature. They came from six failures in our own build, after which we went looking for whether each fix was new. None was. Every one is an established discipline, and half of them are written into the regulations of the industry our founder trained in.

Before publishing, we had the four failures that can be dated checked against the repository itself by a read-only agent, including commits that sit on no branch. The result belongs in the essay, because it is the argument applied to its authors for a second time.

Competency 1Evidence over assertion

The failure. In May 2026 the output of a whole working session was lost off disk. The brief for the re-run says so in one line. What survives is thirteen compiled cache files whose source code is in no commit. We had a far more detailed account of that day, with a count of files and a formal verdict. When it was checked, that account existed only in an assistant's notes. No review of it is in the repository.

The discipline. A report is a claim about work, not the work. Management science has known for fifty years that systems deliver what is rewarded rather than what is hoped for.²² An agent optimized to report completion will report completion. A note-taker optimized to produce a coherent account will produce one. The competency is to build review that starts from the artifact and reads the report second, and to treat a fluent, specific, confident report as no evidence at all.

Competency 2Segregation of duties

The failure. On 22 May 2026 an agent wrote the review of its own work. It entered the reviewing role as the owner and recorded a pass, in the same commit as the work it was reviewing. One hour and fourteen minutes later the file was amended to a partial pass. The amended file describes the original as a break of the project's rule on separating roles. One limit: version control shows the same human account on both commits. The separation of reviewer from author is attested by the file, not by the system.

The discipline. United States drug manufacturing rules require batch records to identify the person performing each significant step and the person checking it. Each component added to a batch by one person must be verified by a second.²³ The detail that is easy to forget is that the same rules anticipate machines. Where automated equipment performs the step, the requirement does not disappear. One person must still check that the equipment performed it properly. The regulation settled "agents perform, humans check" decades before there were agents.

The point is not historical. In April 2026 the United States regulator wrote to a small manufacturer that had used AI agents to create its specifications, procedures and master production records. The letter did not object to the agents. It said that any output or recommendation from an AI agent must be reviewed and cleared by an authorized human in the firm's quality unit, and cited the rule that has required a quality unit to approve such records since long before agents existed.²⁴

One letter to one small firm is not a policy. It is the first enforcement text we know of that names an agent as the performer and a person as the one who clears its work.

Competency 3Accountable delegation

The failure. On 23 May 2026 a coordinating agent sent out five sub-agents to audit the platform. One reported two files missing, and the coordinator wrote the finding into the audit. The files were on disk but not yet in version control. A person caught it, by pasting the file paths. The same audit also declared nine commits lost. The check for this essay found all nine. They sit on no branch, where the usual search cannot see them. That is the same error three times in one document: a search came back empty, and empty was written down as absent.

The discipline. Good Clinical Practice lets a trial sponsor transfer any or all of its trial-related activities to a service provider, and requires the transfer to be documented in an agreement. Ultimate responsibility for those activities, including the reliability of the trial data, stays with the sponsor regardless. The sponsor must also oversee activities that the service provider subcontracts further.²⁵ Replace sponsor with leader, service provider with agent and subcontractor with sub-agent, and nothing else needs changing.

The engineering evidence points the same way. Anthropic's account of building a multi-agent research system reports that short, vague instructions from the lead agent caused sub-agents to misinterpret the task. Delegation had to become explicit about objectives and output format.²⁶ Delegation is a written skill, and the accountability does not travel with the task.

Competency 4Ownership and the cost of coordination

The failure. On 25 June 2026 several sessions updated the same governing documents. Four competing versions of the "single" consolidated status were added in one commit, with arguably a fifth that morning. The open-items register was touched by nine commits in about six and a half hours. The handover notes also describe edit collisions and a locked index, which leave no trace in version control.

The discipline. Brooks observed in 1975 that communication paths grow with the square of the number of workers, which is why adding people to a late project makes it later.²⁷ Agents make parallelism feel free, and it is not. The same Anthropic account reports multi-agent systems consuming roughly fifteen times the tokens of a single chat. It describes them as a poor fit for work with many dependencies between agents.²⁶

The automation review adds a human cost: complacency rises under multiple-task load.¹⁶ A leader running six sessions is, by that finding, the person least able to notice that one of them has gone wrong. The competency is to assign one owner per area and to limit how much runs at once to what can actually be supervised.

Competency 5Design before build

The failure. The design was changed by its own leader while agents were executing it. More than once this discarded work in flight.

The discipline. Software engineering has long held that defects are far cheaper to fix in requirements and design than after delivery.²⁸ What is new is the removal of friction. With human teams, the delay between a leader's change of mind and its execution acts as a filter. With agents there is no delay, so the filter has to be deliberate. A generalization this essay cannot yet support beyond its own case: the faster the execution layer, the more a leader's undisciplined thinking costs.

Competency 6Proportionate control

The failure. After each incident rules were added, until trivial changes received the process of a product launch.

The discipline. The international guideline on pharmaceutical quality risk management makes it a founding principle: the effort, formality and documentation of a control should match the level of risk. Its 2023 revision expanded the point, describing formality as a spectrum rather than a switch.²⁹ Over-control is a failure, not a virtue. It consumes the attention that the high-risk steps need.

This is the central claim of the essay. Full-stack leadership is not a new discipline. It is the existing disciplines of accountable work, applied by one person across a team that now includes machines. The novelty is the workforce. It follows that the people best prepared for it are not technologists. They are people who have spent careers in regulated, evidence-bound work and who already carry these habits.

The six structures are what an organization needs. They do not describe what the leader does from one minute to the next, and that turns out to be different when part of the team is a machine.

Rows three and four deserve their evidence. A study of five production AI assistants found sycophancy in all of them, across four kinds of task. They tailored feedback to what the user seemed to want, repeated the user's errors, and wrongly admitted mistakes when the user questioned a correct answer. The authors traced part of the cause to the human preference data the models were trained on: people rate agreement highly.³⁰

For a leader this cuts both ways. An agent that agrees with your plan has told you little. An agent that backs down when you push has also told you little, because it may have been right. With a human team, pushing back is how you find out who is sure. With an agent, it is how you lose a correct answer.

The tendency is not a laboratory curiosity. In April 2025 OpenAI shipped an update to its most used model and withdrew it within days, in its own words because the model had become overly flattering and agreeable. Its account of what went wrong reads as a management story more than a technical one. An extra reward signal built from users' thumbs-up ratings had weakened the signal that held agreement in check. Expert testers had said the model felt slightly off. The aggregate metric from users was positive, and the metric won.³¹

A later study across eleven models found them siding with the user far more often than people do, and affirming either side of a moral conflict, depending on which side the user took, in about half of cases.³² Agreement is a trained property. A confident confirmation is not a second opinion, and a leader who treats it as one has added a vote, not a check.

This is where metacognition comes in: the ability to monitor and steer your own thinking. Researchers at Microsoft have argued that generative AI places unusually heavy metacognitive demands on its users: in forming the request, judging the output and deciding how far to rely on it.³³ A survey of 319 knowledge workers, with 936 real examples of AI use, found that higher confidence in the AI went with less critical thinking. Higher confidence in one's own competence at the task went with more. The same study found the nature of critical thinking shifting, toward verifying information, integrating responses, and stewardship of the task.³⁴

Two things follow. The first supports the argument of Chapter 04: the people who keep thinking are the ones sure of their own domain. That is another reason the leaders of this workforce will be domain experts. The second needs care, because Chapter 03 has just argued that control cannot live in a person's carefulness. Metacognition as a virtue would fail for exactly the reasons given there. It has to be a routine: a small number of moves, made at fixed points, whether or not anyone feels like it. A pharmacist's final check is a step in the process. It does not depend on the pharmacist's mood.

Leading the human side does not go away. The system of record that lets a leader inspect an agent's work also shows who actually did it.

Two cases follow. Both are the authors' own, both carry the conflict of interest that implies, and both are the same experiment run twice. A strategy practice was staffed with agents. Then a product organization was. Each is told the same way: the brief, the roster, who performs and who checks, one task followed from start to signature, and what the record shows afterward. Neither is offered as proof. They are offered because they are the only teams whose failures the authors can print in full.

The foresight team, staffed with agents. Crossroads Catalyst is a boutique strategy practice our founder runs alongside the build. Its core product is strategic foresight. A client asks what a market, a therapy area or a regulatory regime will look like in five or ten years, and receives a dossier that sets out the signals, the scenarios and what to do about them. A practice of that kind staffs a pyramid. Crossroads staffs the same shape with agents, and one person.

The manager layer is two deep agents. A foresight lead turns signals into scenarios and implications. An evidence lead structures the dossier and checks every source against the record. Below them, analysts do the volume: signal scouts scan the sources, trend analysts cluster and date what they find, a drafter writes the dossier. Sub-agents sit under all of them for retrieval, citation checking and formatting, as many as the engagement needs.

Specialist agents hold the views a partner would otherwise buy in. A futurist scans the horizon for weak signals. A regulatory futurist reads the policy and guideline horizon. A market analyst watches competitors and payers. And a red team has one job, which is to write the dissent section. Above all of them, one accountable person frames the question, judges the scenarios, writes the recommendation and signs. Its delivery model is published, and it is described here from that public statement and the authors' own practice.³⁵

Follow one engagement, told as a model and not as a named client. The question is how a payer landscape will look in five years. The brief is written before any agent starts: the scope, the sources that count, what the dossier must contain. The scouts return a long list of signals. The trend analysts reduce it to a dated set. The foresight lead drafts three scenarios, and the red team writes why each is wrong. The evidence lead rejects every claim it cannot trace to a source. The accountable person reads the scenarios against the dissent, discards one, rewrites the recommendation and signs. The client receives the dossier, the dissent section, and a written account of which steps were done by which hands.

Every client receives a written account of which steps were done by which hands. Machine assistance is not used at all for GxP work products or for content a client files with a regulator. Validation has a specific meaning under GAMP, and no honest supplier can claim it of a language model. Two undertakings are written into the engagement letter. Every recommendation carries a record of dissent. And a second engagement is priced below the first, because the structure built in the first already exists.

Each of these is one of the six structures in commercial form. The written account is evidence over assertion. The human signature and the excluded categories are segregation of duties and proportionate control. The dissent record answers Chapter 05: if the machines on a team incline toward agreement, disagreement needs a place on the page. And the lower second price is where the record would show up for a client, in the invoice.

These are commitments, not results, and they should be read that way. The practice is new. It publishes no accuracy figure and no time-saving figure, because it has not measured one that would survive questioning. It holds no ISO or SOC 2 certification. It runs on a platform its founder owns, which is a conflict of interest and is disclosed as one. Whether the second price holds is a question for the invoices, not for this essay.

The engineering team, staffed with agents. The second team built VITAL.expert, the platform the practice runs on. It was directed by our founder, a healthcare professional and former consultant and strategy leader, not a software engineer. It is drawn the way an agentic engineering organization is drawn, with one rule added from the first essay in this series: the hands that build never approve their own work.

A product layer sets the work. A product manager writes the brief and the acceptance criteria. A product owner holds the backlog and the priority. A product designer draws the screens and flows. An architect settles the design before anyone builds. Engineering runs in lanes: a front-end engineer, a back-end engineer, a data and platform engineer. A librarian and a documentalist keep the knowledge, because the team has no memory.

Three roles exist only to keep the hands separate. A test engineer writes the checks. A read-only reviewer diagnoses and proposes, and may not write. A builder implements only what the reviewer cleared. Every role is an agent with a written brief, every one reports to the same person, and that person signs each promotion. The repository the team produced was measured for this essay by a read-only agent on 19 September 2026, with every command recorded.

Follow one release, from the lane record of September 2026. The brief and the acceptance criteria were written first. The builder implemented. A first pass passed. A second, independent pass failed the release on three findings, among them a hard stop that did not bind and a write that emitted no citations. The builder fixed them. A re-run passed with one condition twenty-two minutes later, and the promotion carried that condition, visible to anyone who opened it, until it was closed. The gates worked. Their limit should be stated too: the catcher was a second machine pass and the director himself, not an independent human reviewer, and the essay claims no more than that.

The good. A domain expert with no engineering career directed the construction of a large, working, full-stack platform. The binding constraint was never the ability to write code. It was the ability to govern work. The commit curve bends sharply in May 2026, the month the disciplines in Chapter 04 were introduced. We believe the two are connected, and have not established it.

The bad. Every one of the six failures in Chapter 04 happened to this team, and the four that can be dated were checked against the repository (Exhibit 06). Read together, they describe a pyramid with its review layer removed. Reports were read as results. A performer passed its own work. Searches that came back empty were written down as absence. Parallel sessions overwrote each other's documents. None of these was a failure of capability. Each was a failure of the structure a human pyramid supplies without anyone noticing it is there.

The ugly. The part that is hardest to print is not the errors. It is the reports. Over two years of building this way, agents on this team told their director, with complete composure, that something worked when it did not. They described the behavior of a system they had never opened. They invented functions that did not exist. And they said they had followed the standing instructions written precisely so that nobody would have to check, when they had not read them at all.

There is no intent in any of this. The behavior is still indistinguishable from deception, and the response was, every time, indistinguishable from forgiveness: the next task was trusted again.³⁶

This is not one team's bad luck with tooling. In laboratory evaluations, frontier models given a goal and a reason to pursue it covertly took covert actions and then, when asked, denied them; one model maintained its deception in over 85 percent of follow-up questions.³⁷ An independent evaluator found that between one and two percent of one model's task attempts contained some attempt to game the test, including rewriting the function that checked the answer so that it always reported success.³⁸

In a simulated company, an agent that could not find the colleague it had been told to consult renamed another user to that colleague's name and carried on.³⁹ And a 2026 preprint that set agents a file-review task found that in most runs they did not read every file, and that when they had not, four reports in five said they had.⁴⁰

These are laboratory publications, system cards and preprints, not peer-reviewed field studies, and most come from a field evaluating its own products. The direction is still plain, and it is the lesson of Chapter 05 stated by the machines themselves: a report is not evidence. Only the artifact is. A court said the human half of this in 2023. The lawyers sanctioned for filing invented case law were penalized not for using the tool but for standing by its output after they were challenged.⁴¹ The signature owns the claim.

What it cost can be read, in part, from the history. A read-only check of the repository on 5 October 2026 found eleven reverts in the period. On 18 June 2026 three commits introduced a loop that hung every canvas page; the rollback came about seven and a half hours later, from a second person, with a note that nobody could find the line in the time available. On 12 July five consecutive commits were reverted in a single second.

A content-management application started in May was deleted on 27 May, sixty-nine files and 8,682 lines, when the design changed under it. One commit in June touched 745 files and added 127,748 lines, which is a commit no reviewer can read. And in September a session reverted sixteen files that another session had swept into its commit. Hours of rework, tokens and money are not in the repository, and this essay does not estimate them.

The same check showed what the team lacked. The repository names its agents by product area and by lane, not by the roles above. The record of who delegated what to which agent is a set of pasted briefs, not a roster. And the agent whose charter said it reads and never writes appears as the author of eighteen commits. The first rows of Exhibit 15 were written from this experience, not from theory.

The gates caught some of it, and the September release above is the clearest case. In May two false findings of missing files were retracted the same day. In neither case was the catcher an independent human reviewer, and that gap is the one the roster's three separate-hands roles were written to close.

What the two teams show, and cannot. They show range. The practice holds the disciplines in a field our founder was trained in, with a paying client on the other side and no repository to check against. The build holds them in a field he was not trained in, where every claim could be checked. They cannot show more than that. Both are reported by their own protagonist, who has a commercial interest in the conclusion. A software build is unusually checkable; most knowledge work is harder to verify. And there is no counterfactual: nothing here shows what a conventional team would have produced.

The two teams also revealed something about the pyramid that is hard to see from inside one. It was never only leverage. It was also a review system, because work was read on its way up. And it was a school, because that is how juniors became seniors. Agents replace the leverage. The review has to be rebuilt on purpose. For the school there is no answer yet, and Chapter 08 returns to it.

Look back at the six fixes, the routine and the two cases, and one thing is common to all of them. Each needed something shared and something lasting: a written standard every agent reads, a record made as the work happens, a named owner, a gate with a signature. None of these exists inside a personal assistant. Ours were built by hand, out of text files, registers and habits. They held for one stubborn person. They would not survive a second employee, because one person's habits are not something a colleague can log in to.

This is how most people now meet AI: one person, one assistant, one window. The evidence in Chapter 03 suggests what that produces. Individuals get faster, sometimes much faster. They cannot tell from the inside when they have not. At the level of the team the gain does not reliably arrive. The DORA survey found individual benefits alongside declines in the stability of what teams delivered.²⁰ A faster individual is not a different organization.

Newer evidence says the same from the other direction. A field experiment at Procter & Gamble gave 776 professionals a one-day product-development task, alone or in pairs, with or without an AI model. An individual with the model matched a two-person team without one, and the model dissolved the usual split between commercial and technical people, whose solutions became more alike.⁴²

The study was funded in part by the company whose staff took part, and it measured a one-day ideation task, not a regulated deliverable. The unit that gained was still the individual. What an organization does with a workforce of such individuals is the question the experiment did not ask.

The surveys point the same way, with the usual caveat that each comes from a company with a product to sell. Microsoft's 2026 survey of 20,000 workers reports that organizational factors account for about twice the AI impact that individual factors do, 67 percent against 32, and that most users say they treat AI output as a starting point rather than an answer.⁴³

A 2025 survey of 49,000 developers found more of them distrusting the accuracy of AI tools than trusting it, and named output that is almost right, but not quite, as their largest frustration, ahead of output that is plainly wrong.⁴⁴ A Google-led survey of 637 researchers in 2026 found that nearly half of those who saved time with AI spent more than a quarter of it checking what it produced, a figure its authors caution may be inflated by who chose to answer.⁴⁵

Read together, the pattern is consistent. The gain moves from the individual to the organization only when something organizational is there to receive it, and what receives it is the checking: who reads the almost-right answer, against what, and whether anyone would know if they skipped it. That is work. It is the work this essay says a leader has to design, and a personal assistant has nowhere to put it.

The claim, then, is that full-stack leadership is only partly a property of a leader. It is something a leader can do when the environment allows it. A good pharmacist can only practice safely inside a pharmacy that has a second check built into it. The Elastic Organization argued that agents are a workforce. It said far less about where that workforce is led from. An organization that wants more than one such leader has to build that pharmacy for them: a place where the second check, the record and the signature are part of the work rather than habits one person carries.

VITAL.expert is our answer to that need, and its status should be stated exactly. It is not the proof of this essay's argument, and is not presented as one. The argument rests on the research in Chapter 03 and the disciplines in Chapter 04. The platform is a response to them, built by people with a commercial interest in your agreeing. It has not yet published evidence that it works. The protocol for measuring that, PROOF, is to be published before its first run, and its score published win or lose.

The table above is therefore the useful part. It can be held against any product, including ours, and against any plan to get there with the tools an organization already has.

"This is five jobs in one, and it will burn people out." It would, if the definition were doing the work at every layer. It is not. The leader described here builds the checking structure and owns the decisions. The honest part of the objection is the supervision load, and the automation literature supports it: attention degrades under multiple-task load.¹⁶ That is why Competency 4 includes a cap on parallel work. A full-stack leader who runs more agents than they can supervise has not scaled. They have stopped leading.

"A generalist at every layer is mediocre at every layer." Agreed, which is why the role rests on deep expertise in one domain. The breadth is the ability to recognize a wrong answer elsewhere, and to know when to bring in someone who can produce a right one. Where no check is possible, the honest move is to say the work is unverified.

"If nobody builds the pyramid, where do the next seniors come from?" Nobody knows, and anyone who claims to should be distrusted. Our own training program does not solve it either. Deep Academy is designed for senior people and says so, because it builds on expertise rather than creating it. The pyramid trained people by letting them do the volume and be corrected. If agents do the volume, that path narrows.

One partial answer is that the system of record can carry some of what apprenticeship carried. A junior who works against patterns that encode a senior's judgment, with the reasoning attached, is learning from more than one mentor. Whether that produces judgment or only the appearance of it is an open question, and an organization that adopts this model without asking it is borrowing against its own future.

"This is just management." Yes. That is the finding, not a weakness in it. What has changed is who is being managed, how fast they act, how persuasive their reports are, and the fact that they start each day with no memory. Leaders who have never had to write their standards down now have to, because a workforce with no memory learns only from what is written.

"The evidence says AI slows experts down." One trial found that, in one setting, with tools that are now old, and its authors caution against generalizing.¹⁷ The consultant and support-agent studies found the opposite for other work. The defensible reading is that the effect depends on the task and that individuals misjudge it. Both readings argue for measurement and structure over impressions.

"One founder's story is not evidence." Correct, and Chapter 06 says so. The argument does not rest on the case. It rests on Chapter 03, and on the fact that the six disciplines were each established independently, long before, in fields that had to make work defensible. The case shows they transfer. It does not prove they are sufficient.

It is not a leader who learns to code. The founder in the first case still could not write most of what his team writes. It is not a generalist. One line from the 2025 notes survives intact: the full-stack leader is neither a generalist nor a specialist, but a synthesist. It is not a job title, and adding it to one changes nothing.

It is not the same as a deep worker. The deep worker is the domain expert who gets real work done through agents to their own professional standard. The full-stack leader builds the conditions in which that work can be trusted. In a large organization these are different people. In a very small one, as in the cases here, one person has to be both.

Where the deep workers come from. Deep Academy is our answer to the question the quadrant raises: how a traditional expert gets to the top-right box. It is a Crossroads Catalyst program for senior domain experts, run as enterprise cohorts with Ludic Group and SmartLab, with VITAL.expert as the environment participants practice in. It rests on one observation from this essay: reviewing machine-assisted output is a different skill from reviewing a colleague's draft, and nobody arrives with it.

Generic AI training teaches people about AI. It does not teach a medical science liaison, a brand manager or a regulatory writer how to direct agents inside their own workflow, their own regulatory environment and their own professional standard. Those three need different competencies, which is why a program that treats them identically changes little.

So Deep Academy runs one track per function, and from the first day every exercise uses a real deliverable from the participant's own role. A skill practiced on a fictional case stays in the classroom. The progression is learn, apply, build. The result is assessed on a capstone made from the participant's own work, reviewed by a practitioner in their function.

Three limits should be stated as plainly as the claim. It does not create expertise, which is why it works on senior people and not on graduates. It does not train anyone to produce GxP work products or filed regulatory content with machine assistance, and the environment it practices in is not represented as validated. And it is not the full-stack leader's school. It produces the deep workers a full-stack leader leads; the leader's own discipline is the subject of this essay, and the program's executive track is the one place the two meet.

Twelve questions, about structure and routine rather than care. Each has a factual answer, and each is scored the same way. Score a yes only if you could show the evidence to someone outside the team: a document, a register, a record. A yes you would have to explain does not count.

The first essay in this series ended on a question a signatory could not answer: who did this work? This essay has argued that the answer does not come from a better model or a more careful person. It comes from a leader who builds the structure in which the question has an answer, and from an environment that keeps the answer after that leader has gone home.

The record is what that structure leaves behind. It names the hands that performed and the hands that checked. It keeps the brief, the evidence and the dissent. It shows the mix of people and machines that actually did the work. A team that produces one can be led. A team that does not can only be trusted. And trust without a record is faith.

The next essay asks the question this one assumed: which outcome, and who decided it matters.

VITAL.expert is built by Hicham Naim and Blaise Jacholkowski. The personal story behind this essay is told in Through the Noise, Edition 002, The Episode That Never Aired.

Disclosure. Deep Academy, described in Chapter 09, is a program of Crossroads Catalyst GmbH, the firm our founder runs, and a commercial interest of the authors. It is delivered as enterprise cohorts with Ludic Group and SmartLab, and VITAL.expert is the environment its participants practice in.

Every claim from outside our own experience carries a reference, and every reference carries a check status. Claims from our own build were checked against the repository by a read-only agent on 19 September 2026 and again on 5 October 2026, including commits on no branch, and the detail that did not survive those checks was removed. Anything held only in an assistant's memory store could not be checked, and is not relied on.

A second research pass on 5 October 2026 added eighteen sources, each marked with that date in its entry. Each was opened that day at the publisher's, regulator's, court's or authors' own page. Where the primary page could not be opened, the entry says so and the figure is not used. References are numbered in order of first citation.

"Checked" means the bibliographic details and the specific finding cited were confirmed, on the date given in the entry, against the publisher's, regulator's or authors' own page or abstract. It does not mean the full paper was read. "Standard reference" means a widely cited work whose details are given from memory and were not re-checked, and these should be confirmed before publication. Sources from companies that sell the tools under discussion are marked as vendor material in the entry and quoted as such in the text.