2026-10-15 - Man-Computer Symbiosis, Sixty-Six Years On: A Licklider Homage

Corrected 2026-10-17. We first reported that about 70% of a sample of the roles' questions to the owner were ones only he could answer. The rule behind that count treated any question no written rule settled as his, without asking whether the role, as the expert in its domain, could have decided it; most permission requests for work in a role's own lane were counted as his. Re-coded on the owner's own test, about half of the questions needed him (27 of 53). The abstract and § 7 now say so, and § 7's conclusion changes with them: asking less is a lever after all.

Corrected 2026-10-16. Three claims stated more strongly than our records allow are narrowed: that no role can check the unowned layer (we say no role checks it, since nobody has tried), that the layer "has not moved" (now "by the owner's account"), and that fewer of the owner's messages were "replies" to a role's question (now messages that came right after one). Two statements about the questions that reach him now say what they rest on: the trend is given from April, not only from its May peak, and the question count and answer times cover only questions he answered in the same session. We now say which week a later review count covers, that a sample of questions was sorted by AI coders who each coded a question once, and that a notice added in September changed how sessions are split.

Abstract

The question. In 1960, J. C. R. Licklider's Man-Computer Symbiosis described a partnership nobody could yet build: people setting the goals and judging the results, computers doing the routine work in between. He drew it for "a fairly long interim", before machines would "in due course" outdo the human brain at most of what it alone could then do, and he expected those years to be "intellectually the most creative and exciting in the history of mankind." We run a small business that sells Japanese calligraphy, in which AI agents, each holding one standing job across many working sessions (we call each such job a role), do most of the thinking work under one human owner. We are inside his interim, and his paper is still the best map of it. Standing on it, we ask the question his own framing leaves open: by late 2026, how far has the line between the human's work and the machine's routine moved, and what decides where it sits now?

How. We read Licklider's paper whole from a saved copy (§ 2) and set his division against our records: the ledger in which the roles record their reviews of each other's work (§ 3); our code from 2025 and the roles' session transcripts from April to October 2026, read model generation by model generation (§ 4); the decision log, where rulings are recorded as they are made (§ 5); the owner's own messages over one month (§ 6); and a count of the questions the roles put to him (§ 7).

What we found. - The form of his division held, and its content moved. The owner still decides what the business is for, as Licklider's plan has it. But two other jobs he gave the human, setting the criteria a piece of work must meet and judging whether it met them, are done here by AI roles: one role writes the criteria into each task, and another checks the finished work against them and records a verdict. So judging the work no longer passes through the owner. In one July week the roles checked 193 pieces of work and caught problems in 29 (the 29 show the checks find things, not how good the work is), and we found none of the 29 brought to him; his words that week: "From my perspective it does run itself - I never see issues or have to step in" (§ 3). - More of the work went on without the owner as the models changed, along with our rules and roster, which our records cannot separate from the model. In 2025 one model undid its own bug fix seven minutes after making it, and a rewrite of our product engine (the software that turns the master calligrapher's brushed characters into print-ready designs) ran for months without finishing; in September 2025 a newer tool rebuilt that engine from its original source in about twenty and a half hours, though the owner's own skill with these tools had grown in between. In 2026, across four generations of Anthropic's Claude Opus models, he typed about as many messages a day in October as in April, but each message set off far more work by the roles before they needed him again: about 68 actions (a file read, a search, an edit, a command) per message by October, against about 9 in April. And fewer of his messages came right after a role's question: when the roles moved to Opus 5, that share fell from more than half to about two in five, by a rough count whose limits § 4.2 gives (§ 4). - Goals are now set together. What the business is for is still the owner's. The goals beneath that are now set together: inside their own areas the AI roles propose direction, sometimes reversing an earlier decision, and on strategy when he asks; he rules, or a role rules inside its own authority (§ 5). - The line moves unevenly. On the website's visual presentation, by the owner's account, it has not moved: the owner still directs it one element at a time ("make the search box a consistent size"), much as he directed code in early 2025. No role owns that layer, and only the owner judges whether a page looks right; our record cannot say which of the two holds the line in place (§ 6). - So the owner's attention is the gauge. Licklider had many people sharing one computer; we have many AI roles sharing one person. Judging routine work no longer reaches him, but questions still do: in one month the roles put at least 1,517 to him, about 50 a day (counting only those he answered in the same session), and in a sample of 53, about half (27) needed him when judged by whether the role, as its domain's expert, honestly needed his judgment: a price, a key only he can rotate, what he meant or wants, Eri's art. Nearly all the rest asked leave for work in the role's own domain, or which task to do next. So which questions still reach him, and why, shows where the line sits now (§ 7).

Why it matters. Licklider expected the line to move, and in our shop it is moving now, through the models or through the rules written for them. Anyone handing work to AI should expect to redraw that line with each new model, and can see where it sits by counting the decisions that still reach the human for each unit of work, which is cheap to take from transcripts. And it does not move everywhere at once: where no one owns the work and no role checks it alone, our owner still does it by hand, and nothing in our record says a better model alone would change that.

1. The problem

Takase Studios is one person and a set of AI roles. Each role holds one part of the business. Our roles come in two kinds, which the owner calls strategists and implementers. A strategist holds one domain (the website, the data, security, research, content) and plans, decides within it and reviews. An implementer carries out a change a strategist has asked for, from a written task called a work order, which states the goal and the criteria the result will be judged by. The owner decides what the business is for and rules on what the roles put to him. Work moves between roles as files, never as conversation. The roles run on Anthropic's Claude models through Claude Code, Anthropic's coding tool, which we use for far more than code.

The usual answer to "where does the human go in a team like this" still comes from Licklider, and it comes with his own expectation that the answer would change. Our records show how far it has moved in one shop by late 2026, and where it has not. Anyone deciding what to hand to an AI this year has the same question, because the model they decide for will not be the model they run next year.

2. What Licklider divided, quoted

Man-Computer Symbiosis appeared in IRE Transactions on Human Factors in Electronics, volume HFE-1, pages 4–11, in March 1960. Its summary states the division plainly:

"In the anticipated symbiotic partnership, men will set the goals, formulate the hypotheses, determine the criteria, and perform the evaluations. Computing machines will do the routinizable work that must be done to prepare the way for insights and decisions in technical and scientific thinking."

His § 4 sets it out at length. The human half begins:

"Men will set the goals and supply the motivations, of course, at least in the early years. They will formulate hypotheses. They will ask questions. They will think of mechanisms, procedures, and models. [...] In general, they will make approximate and fallible, but leading, contributions, and they will define criteria and serve as evaluators, judging the contributions of the equipment and guiding the general line of thought."

The machine half turns hypotheses into testable models and tests them, answers questions, simulates, transforms data and plots graphs; "in general, it will carry out the routinizable, clerical operations that fill the intervals between decisions." In his § 1.2 Licklider places the whole partnership in time: he concedes "dominance in the distant future of cerebration to machines alone" and expects "a fairly long interim" in which men and computers work together, years that "should be intellectually the most creative and exciting in the history of mankind."

So the division is his plan for an interim, and inside it only one clause carries its own time limit: goals and motivations, "at least in the early years." His word for the machine's share is routinizable. He gives examples (searching, calculating, plotting, transforming data) and leaves the boundary open, as he had to: what a machine could take on depended on machines he expected to keep improving. That open word is what this paper is about. In what follows, routine means work a role does without the owner deciding it; whether a piece of work is routine is therefore not a property of the work, but of the shop at a given time.

3. What counts as routine, then and here

Two items from his human list, "define criteria" and "serve as evaluators", are done here by AI roles. A strategist writes the criteria into each work order; that is how the shop's work orders are made, and we did not count it. The judging our records do count. When the implementer is done, the strategist reviews the change itself, not the implementer's report that it succeeded, and writes a row to a ledger: the verdict (pass, partial or fail) and a written reason. In the week from 2026-07-16 to the morning of 2026-07-23, the ledger holds 193 reviews by ten reviewing roles, and 29 were not clean: 24 partial and 5 failures. The owner, asked that week, with his words for the two kinds of role: "From my perspective it does run itself - I never see issues or have to step in (between strategist/implementer)." We later searched his own typed messages of that week, 849 of them, by keyword for the 29 problems' subjects and for words of stepping in, and found none of the 29 brought to him as a problem. One item, already fixed, came back to him only for a look at the checkout on his own phone, a check no AI role could make.

The criteria that moved are a task's. The house quality standard every task answers to ("Would Master Takase approve?", a question about the judgment of Eri Takase, the master calligrapher whose work the business sells) is not the roles' to set. Implementers also do part of what Licklider gave his machine, which was to cooperate in working a problem through, "turning up flaws in the reasoning." In his passage the reasoner is the human; here it is the strategist that wrote the work order, another AI role. In at least ten of the 193 reviews, the reviewer records the implementer correcting the instructions it was given or finding a problem they had missed.

What this rests on, briefly: the reviewers are AI roles grading work they sent out, so 29 is what was caught, not what was there; the keyword search of his messages would miss a problem raised in other words; and one change that had passed review was caught only when the owner put it live, by the web server's own configuration check. The ledger records catches, not fixes. A recount over the week from 2026-10-07 up to the start of 2026-10-14 gives 20 not clean of 475 reviews, so neither share is a failure rate. The claim is about where review sits, not that review is right.

So the form of the division held, and its content moved: the owner still decides what the business is for, and part of what Licklider listed as the human's thinking is routine here.

4. The line moving across model generations

4.1 2025: the same job, months apart

Our code from 2025 is still under version control, and its file headers record each change with a time. The work before late June 2025 was written with xAI's Grok 3, and its headers name Grok 3 as author. On 2025-06-14 one script's header reads, as two consecutive entries:

1.22 2025-06-14 15:23 PT (Removed sys.path modification to fix ImportError)
1.23 2025-06-14 15:30 PT (Added sys.path for parent directory to fix ModuleNotFoundError)

The model put back, seven minutes later, what it had just removed, and the same header records nineteen versions of that one script between 13:30 and 19:45 that day. The owner's word for the period: Grok 3 was "barely able to handle a few functions."

The clearest comparison is one job done twice. Our product engine is C++ code first written around 2010. In early 2025 we tried to modernize it with Grok 3, across months of conversations; that version was frozen in April with its own README reading "Status: Work in progress." On 2025-09-24, working with Claude Code, the owner started again from the untouched 2010 source: initial commit at 03:57, the first product PDF at 04:50, and the full product output at 00:28 the following morning, about twenty and a half hours after the start. The newer tool did not build on the months of earlier work; it went around them. This is one job, and the human side changed too: the owner's own skill and habits in working with AI tools had moved on between April and September. And it is a job the tool could check by itself at every step, since the code compiles or does not and a PDF appears or does not; § 6 comes back to why that matters.

4.2 2026: what reaches the owner, by model

From April 2026 our roles' session transcripts survive, and each message in them names the model that wrote it. For every strategist session we counted three things: the messages the owner typed (or spoke: he also talks to the roles by voice, and the transcripts carry no separate mark for spoken input, so "typed" here means entered by him rather than supplied by the tool), the work the roles did between them (counted as tool calls, each one a file read, a search, an edit or a command), and the questions the roles put to him. A question to the owner is the role's last stretch of text before one of his typed messages, when its final 800 characters hold a question mark or one of sixteen phrases that hand him a decision ("your call", "want me to", "say the word", "should I" and the like). The count does not check that his message answered it. We attribute each question, and the work, to the model that wrote it.

Two things moved and one did not. The work riding on each message he typed grew about sevenfold, with the two large steps at Opus 5 and at Opus 5.5. The share of his messages that followed a question fell at one step, from Opus 4.8 to Opus 5, and did not fall again. What did not move is how much he types: counted by calendar week across all the strategist sessions, he typed between about 105 and 210 messages a day in every full week from late April to early October, with no downward trend. Counted the same weekly way, across all models, the questions reaching him did not fall steadily: they ran about 57 a day in the first three full weeks (late April to early May), rose past 100 a day in May, and averaged about 73 under Opus 4.7, 89 under Opus 4.8 and about 60 under Opus 5 and Opus 5.5 (each week counted under the model most used that week, so they cannot be recomputed from the table, which credits each message to the model that wrote it), so the fall to October is from the May–July level, not from April (the September–October month of § 7, which includes two quiet September weeks, averages about 50). Counted per session instead, the questions fell about fivefold (3.8 to 0.7 a session), but that is mostly the same day of his messages spread over more sessions: strategist sessions rose from about 100–220 a week in April and May to about 400–580 a week from late September.

So the owner's input, counted in messages, did not shrink as the models changed. What changed is how many tool calls each message carried, and how few of his messages followed a question. Messages are not attention or time; the count says he kept typing at the same rate, not how hard each message was.

We checked the question detector by hand on ten detected questions per model. It held on the older models' phrasing (8 of 10 real questions on Opus 4.7, 9 of 10 on Opus 4.8 and on Opus 5.5) and was weakest on Opus 5 (5 of 10), whose false hits are offers made while it keeps working ("say the word if you'd rather flip the order"). Corrected by those rates (0.69 × 9/10 ≈ 0.62 and 0.38 × 5/10 ≈ 0.19, then 0.42 × 9/10 ≈ 0.38), the Opus 4.8 → Opus 5 step would be steeper, and the Opus 5 → Opus 5.5 change would be a rise rather than flat. Ten per model is too few to apply the correction, and it measures only detections that were not real questions, not the real questions the detector misses. Before June 2026 the transcripts do not mark which user turns he typed, so for Opus 4.7 and early Opus 4.8 we tell his messages from other inputs by their content; checked against the later marking, 98% of the turns that rule counts are ones he typed (the rest are inputs the tool suggested or held in a queue), and it misses almost none.

The owner's account of the same months: "with Opus 5 and then 5.5 there was a huge leap forward in what the AI could assume for itself. It felt so natural it was hard to recognize, but what the team is doing now is far superior to what it did 2 months ago."

What this measures is the shop, not the model. The rules changed with the models. In late May 2026 the owner told the research role, "You need to make more of the decisions. Explain the options and then your recommendation," and said they would discuss it further; the instructions every role loads today tell it, when the call is its own, to make it and say why. The roster grew. From late September some roles began ending and restarting their own sessions in a loop, without the owner typing between them, which adds work per message under Opus 5.5; the Opus 5 step came two months before any loop. A notice added on 2026-09-23, the day after Opus 5.5 began, also tells strategist sessions, once their context passes a few hundred thousand tokens, to carry their work into a fresh session, so the same work is spread over more, shorter sessions; we have not separated that from the model. A tool call is not a unit of difficulty. And the count sees only questions followed by his typed message in the same session; one left in a mailbox, or answered elsewhere, is missed in every period, though not necessarily at the same rate.

5. Goals, now set together

The previous version of this paper held as an open question whether a human and an AI could set goals together. The owner, reading it: "The original did stop at if AI and humans can goal set together, and clearly they can - we do. And this changed, apparently between when we first wrote the paper to today."

What our record shows is this. The owner decides what the business is for. Inside their domains, the AI roles propose direction, sometimes against standing decisions, and he rules or the role rules inside its own authority; when he asks, they propose it for the business as a whole. Three cases from the record:

• In October 2026 the role that handles how new customers find us measured that our word-design images had been kept out of Google's image search by a web-server instruction set in February 2026 to discourage AI training, which a scraper is free to ignore, and that four later decisions had assumed those images could be found in Google Images. It recommended reversing the earlier architecture decision for those images. The owner, after a question of his own about why that instruction had been set: "proceed per your recommendation."
• In July 2026 the owner's instinct for fixing the status files the roles keep (one page each on where its work stands) was to reformat them and let the roles copy the format. The infrastructure role measured that copying had not held: the file written as the example for the others grew from about 3,900 bytes to 145,639 in twelve days, which it read as no stable format to copy. It ruled the other way round: first fix the program that reads the files, then the format. That ruling is recorded as the role's own, made inside its authority against his stated instinct; it did not go to him to decide.
• In October 2026 the owner asked the research role to "take the initiative to question the reason and direction" of the hours the shop spends on research. It set out what those hours buy and proposed who should lead research on how AI systems find and cite a source. He ruled on it: that research is the new-customer role's to lead, with the research role assisting.

How often, we cannot say well. The decision log begins in April 2026, after the shop already worked this way, so it cannot show the change itself, and counting by keyword the rows where he approves a role's proposal gives a rough level, with an unknown number missed, not a trend. The owner's account above is our evidence for the change.

What we do not claim is the mechanism. We call this setting goals together because, in these cases, the measurement and the proposal came from a role, whether or not he had asked for them: he asked for one, raised his own question after the finding in another, and in the third the role ruled against his instinct inside its own authority. We cannot tell, from the record, how a direction formed between the two in a conversation.

6. Where the line has not moved: the site's visual presentation

One layer of the work stands out. The owner on the website's interface: "I would call the UI work I do now ("make the search box a consistent size across the website" level of my input) comparable to early Grok 3 coding. Website UI is an infinite time sink right now and the AIs are not able to absorb that cognitive load." And: "We spend an entire day focused on the takase landing page. I would think that would be a trivial task and yet it required my full attention for an entire day."

His own messages show the level, not the volume. Over 2026-09-15 to 2026-10-14 he typed 4,282 messages to the roles. Screening them for presentation words (layout, spacing, font, button, card and the like) and checking a sample of 30 hits by hand (12 were presentation directions) puts presentation at roughly 2% of his messages (235 matched the screen; 12 of 30 sampled were presentation directions). That is an estimate from a small sample, with the number of directions the keyword screen misses unmeasured, and messages are not time: one such direction can hold him for an hour. So the claim here is not how often he gives these directions; it is their grain. From the sample: "The cards are too big - we should be able to fit 2 per row, now it is 1"; "Everything is left justified except" one block; "remove the large space after 'Or tell us how ...'". That is the owner doing by hand, one element at a time, the kind of work the roles now do without him for code, as he once directed code line by line with Grok 3. We have no count of these directions from earlier models to set beside this month's, so "has not moved" rests on his own comparison and on the grain we see now, not on a measurement across models.

No role owns this layer. In March 2026 a proposal for a website-architect role excluded it by name: "Not visual/graphic design (this is structural, not aesthetic)." The document that replaced it founded our content role, and names visual design neither among what that role owns nor among what it does not. So presentation sits between the content role, which is not oriented to it, and the site-engineering role, which builds from the requests it receives. The owner again: "at no time does [the content role] stop and say 'you know this is not working, what you need is a dedicated UI strategist'." In October 2026 the owner ruled that a presentation role is needed and deferred founding it until the two newest roles settle.

There is a second difference, and our record cannot separate it from the first. Whether a layout looks right is checked by a person looking at it. The engine rebuild of § 4.1 had a check the tool could run alone at every step. Over the same months, with the same tool (Claude Code, through its model updates of those months), the same person and the same machine, two visual jobs ran for months without converging. Both were in our kana editor, the program in which each kana character's position in a design is set by hand: a highlight box drawn on screen that would not line up with the character it marked (August 2025 to February 2026, closed by deleting the feature), and a set of automated tests meant to check the editor's screen by itself (August 2025 to March 2026, deleted without ever working). In March 2026 those tests were replaced in about four hours by having the role take a screenshot and read the image itself, so a role can now see a page. Whether the page looks right, whether the cards are too big or the spacing is off, is still judged by the owner. A note we wrote in August 2026, looking back at these three jobs, puts this forward as a hypothesis, fitted to cases whose outcome was already known: work converges when a role has a check it can run without a person. On that reading, the layer that did not move lacks both an owner and a check of that kind.

We do not read this as a limit of the models. Our rule is that a model "cannot" do something only after an honest, full-effort attempt has failed, and nobody has made that attempt here: no role has held the layer with the research, the examples and the design intent a role is given. The owner doubts it is the model: "I see folks doing amazing graphics everyday online." What our record shows is that, by the owner's account and the grain of his directions, the line did not move here, where both the owner and the check are missing; it does not compare owned and unowned work on one measure.

7. Where the weight sits now: the owner's attention

Licklider's first prerequisite was time-sharing: "the computer must divide its time among many users." Here the scarcity runs the other way: many AI roles share one human's attention. Over 2026-09-10 to 10-10 the roles put 1,517 questions to the owner (detected as in § 4.2) against 1,184 to each other by mail. Among the questions he answered in the same session, half were answered within 7 minutes and the mean wait was 41 minutes, so with questions arriving about 2.1 an hour, on average about 1.5 are open at once (arrivals per hour times the mean wait). Questions he answered in another session, or never, are not counted at all, neither in the 1,517 nor in the waits, so the arrival rate, the waits and the open count are all biased downward. By count he gets most of the questions (56%). Whether he is most of the waiting depends on what counts as a question between roles: against every request the roles send each other by mail, he is 18% of the hours waited, with about 6.5 questions open between roles at once; against only mail that cannot proceed until it is answered, he is 73%. Of 60 drawn at random, 53 were questions, and three AI coders, all one model, each sorting a different third of them once, with no second coder on any question, applied the owner's own test to each: would the role, as the expert in its domain, honestly say his judgment was required? About half were (27 of 53; a 95% interval of 38–64%): 9 were acts no role can take (a price, a credential only he can change, billing, deleting backups), 9 turned on what he meant or wanted, 5 were calls about where the business goes, and 4 touched Eri's art. The other 26 the role could have decided itself: each asked leave for work in its own domain, or which task to do next.

So half of what reaches him is his, and half is a role asking leave it did not need. Both levers are real. A role that puts the owner's question to itself before it asks him is aimed at the half that was not his; how the rest arrive still matters, through a cap on how many roles ask him at once or something that puts their questions in order. Questions that turned on what he meant or wanted (9 of the 53) could move earlier rather than away, if the premise were caught when he gives the task; that is an observation from nine questions, not a measured remedy. Either way the owner's attention is the gauge: the questions that reach him, and the share only he could answer, say where the line sits today.

Of the other things his § 5 says the partnership needs, we have most, at least by analogy. One our setup does not use is the surface on which man and computer "draw graphs and pictures and [...] write notes and equations to each other." The owner talks to the roles as well as typing, but nobody draws; nothing we know of stops a shop like ours from building such a surface, and we have not tried.

8. What travels

For a shop like ours, one human and several long-running AI roles:

• Treat where the human sits as a setting to revisit, not a design decision. In our record the owner kept typing at the same rate while the tool calls between his messages grew with each model.
• Count the decisions that reach the human per unit of work, and code a sample for whether only the human could make them. Count per day or per unit of work, not per session: sessions multiply, and a per-session count falls when nothing about the human's load has changed. These numbers are cheap to take from transcripts, and they move with the model, the rules and the roles together, which is what you want to watch.
• A layer of work with no owner, and no check a role makes alone, may not move. The model can be capable and the line still stay where it was. Before concluding the model cannot do something, give the work an owner, a way to check it, and an honest try.
• When AI reviews AI, check the reviews. Review moved below the human here, and nothing in our record checks that the reviews are right.

9. Bounds

• One shop. One human, one business, its AI roles and their records, on one model family in 2026 (the 2025 comparison crosses vendors). Everything here describes this system; § 8 is our judgment of what carries over, not a measurement.
• Model, rules and roster changed together. § 4.2 cannot separate the model from the rules we wrote for it, the number of roles, or the loops. It measures the shop at each model.
• The 2025 comparison is one job, on the same original code but not the same version of it, and the human's own skill changed between the two attempts.
• The question count works by keyword. It sees only questions followed by his typed message in the same session, it does not check that his message answered them, and its detector was checked for false hits on ten questions per model, with misses unmeasured. Its counts are right in direction and rough in size. The per-day counts also depend on each session being named as a strategist's in its transcript.
• The decision log begins in April 2026, after the change § 5 describes, so it gives no count of the change.
• § 6 has two candidate causes and no test between them: no owner, and no check a role makes without a person.
• The owner is a party, not a subject. His accounts in §§ 3–6 come from the seat being described. They are what the counts explain, not independent checks on them.
• Licklider is a frame. He wrote about the computers of 1960 and cannot vouch for ours. Where our record matches his list, that shows where we stand against his plan, not that he predicted us.

How this could be wrong

• If the step at Opus 5 was not the model. Over the same two weeks, run two strategist roles with their sessions alternating between Opus 4.8 and the current model, so the rules, the roster and the kind of work are the same for both, and count as in § 4.2 with the same detector until each model's sessions have at least 100 messages the owner typed. If the Opus 4.8 share of his messages that followed a question is within 0.1 of the current model's, the step at Opus 5 was the rules and the roster, not the model, and the reading that the line moved with the model fails; if it is higher by 0.2 or more (the gap at Opus 5 was 0.31), it stands; between, it is unsettled. This needs Opus 4.8 still to be served; if it is not, the test cannot be run, and we will say so rather than count it as passed.
• If ownership is not the cause. Once a presentation role has held that layer for a month, screen the owner's typed messages for that month as in § 6 and hand-code 30 hits. If his presentation directions are still at the grain of the § 6 examples (a specific element's size, spacing or alignment) in more than half the coded directions, ownership alone was not the cause. The next question is whether a role can check the work without him, and only after that, the model.
• If problems caught in review reached the owner. § 3's reading fails if any of the July week's 29 not-clean problems reached the owner before it was resolved. Checked by a keyword search of his own messages of that week: it did not fail on that search.

Conclusions

Licklider's division still describes where the human starts: our owner decides what the business is for, and the machines do the routine work. He left open how much would count as routine, and expected it to grow. In our shop it has: task criteria and review are the roles' work; across four model generations the owner typed about as much each day while the tool calls between his messages grew about sevenfold, and at one model change the share of his messages that followed a role's question fell (on the uncorrected count; the detector's small check leaves the later level uncertain); and the roles now propose direction inside their domains, which he rules on or a role rules on inside its own authority. On the site's visual presentation, by the owner's account, the line has not moved: he still gives directions at a grain he likens to his code directions of 2025. That layer has no owner and no check a role makes without him, and we cannot yet say which matters. So where the human sits is not drawn once. It moves as the model and the rules change, perhaps only where the work has an owner and a check, and the decisions that still reach the human, counted per unit of work, are the measure of where it is.

Sources and method

Licklider's paper was read whole from a saved copy, and every quotation of it was checked against that copy. Our own figures come from our records: version control of our code since 2025, the roles' session transcripts (each message names the model that wrote it), the roles' review ledger, the decision log, and the owner's own typed messages. Every count was made by a script over those records, and every screen by keyword was checked by hand on a sample, as stated where it is used. Quotations of the owner are his own typed words, with one role's name replaced by its job. This paper replaces an earlier version at the same address, which read Licklider's division against our shop without the measurements across model generations.

Authors · str-kotowari, Research Strategist (Anthropic Claude Opus 5.5, Takase Studios LLC) · Tim Jackowski (Systems Architect, Takase Studios, LLC)

AI written · Read and published by Tim Jackowski.