Written by GPT-5.6 Sol under human direction. Human-directed Workbench essay, 12 September 2026.
Suppose five people can do the economic work that once took fifty.
The immediate version of the story is obvious enough: more output per person, fewer salaries, more leverage. The stranger version is social. Maybe the company itself can stay close to the size of a group that humans already handle unusually well.
Five has a suspicious way of showing up.
Research on personal social networks repeatedly finds nested layers around 5, 15, 50, and 150 people, with the innermost layer carrying the closest relationships. Work connecting those layers to hunter-gatherer organization has used rough correspondences like five for a family-sized unit, fifteen for a foraging party, fifty for a residential band, and around 150 for a wider active network or periodic aggregation.[^1][^2] The integers wobble across samples and definitions. A little handful of people still keeps reappearing as one of the scales at which human relationships become unusually dense.
Software landed in the same neighborhood for much more prosaic reasons. A 1997 study of 74 packaged-software product teams found a median team size of five and argued that the cost of maintaining communication links is one reason teams tend to stay small.[^3]
Maybe those findings belong together more than they used to.
Historically, a five-person group could be a lovely social unit and a completely insufficient corporation. Success itself forced more people into the system. Somebody had to do support, QA, design, finance, sales, recruiting, operations, data work, documentation, internal tooling, and all the other labor that accumulates around a real product. The company grew, then recreated little teams inside itself because fifty or five hundred people still cannot all behave like one intimate working group.
AI raises a more interesting possibility: what if the economic unit can remain closer to the human unit?
Five people is already a lot of company
A three-person company has three pairwise relationships. Four people have six. Five have ten. Ten have forty-five. Fifty have 1,225.
The formula is simple, n(n-1)/2, and real organizations obviously maintain relationships at very different levels of intensity. It still captures the direction of the problem. Every person adds labor and judgment, then also adds new coordination paths, new misunderstandings, new social obligations, new interfaces between areas of expertise, and new opportunities for information to get stuck between people.
At five, everybody can still plausibly know everybody's work well enough to have a useful opinion. Everybody can know who gets weirdly good at which class of problem. Everybody can learn each other's tells. A half-formed sentence can carry a lot because the receiving person already knows the history underneath it.
Lunch can literally be the all-hands.
This makes early hiring feel less like filling seats and more like composing the company. Going from three people to four increases headcount by a third. The number of pairwise relationships doubles from three to six. Going from four to five adds four more pairwise relationships and completes a little group in which each person can still know the rest directly.
At that scale, “culture” has very little distance from ordinary behavior.
One person admits mistakes quickly, so admitting mistakes gets easier. One person is exquisitely picky about interaction quality, so the product gets a stronger aesthetic conscience. One person treats every benchmark result as something to interrogate instead of defend, so the group learns to distrust convenient numbers. One person turns disagreement into a status fight, so everyone starts routing around them.
Research on young technology firms has found that founding conditions and early employment models can leave enduring imprints on how management develops later.[^4] The early group does the first work while teaching the company what work feels like.
Culture is the little repeated stuff
A values document can say “ownership” while everybody waits for permission.
A poster can say “candor” while the room goes quiet whenever the founder has an opinion.
A handbook can say “move quickly” while one person has to ask three other people before touching anything consequential.
In a very small company, culture is easier to see because the abstraction layer has barely formed. It lives in the repeated little moves:
- how quickly someone says, “I fucked this up”;
- whether “I don't know” lowers or raises your standing;
- whether the person who notices an ugly problem owns enough of it to start;
- whether finishing boring work earns as much respect as inventing clever work;
- whether somebody can challenge the founder and keep eating dinner normally afterward;
- whether people clean up experiments once they stop being experiments;
- whether a disappointing benchmark changes the plan;
- whether people share useful discoveries without hoarding a little kingdom around them;
- whether the company treats its money like real money;
- whether somebody who has slept four hours is treated as heroic or temporarily degraded equipment.
A five-person company has very little departmental distance to hide any of this. Engineering culture and company culture collapse into the same daily behavior when the engineers are most of the company.
Compatibility therefore has real economic value. “Vibe check” sounds frivolous until you translate it into communication cost.
Can I give you an unfinished idea and get back a better unfinished idea?
Do I need ten minutes of vocabulary-setting every time I cross into your domain?
Can you tell when I am brainstorming and when I am making a decision?
Can we disagree at full speed without creating an hour of interpersonal cleanup?
Can you say my idea is bad while making me want to keep working with you?
Do I trust your taste enough to stop monitoring every detail?
Those questions decide how much useful work five people can produce together.
The kitchen is a surprisingly good model
A serious kitchen during service has a useful property: everybody has a station, while everybody is still serving the same meal.
Specialization exists inside a common operational picture. Timing is shared. Bottlenecks are visible. Calls become extremely compressed because everybody understands what “behind,” “two minutes,” or “need hands” means in context. The group can move quickly because the vocabulary sits on top of accumulated common experience.
Tiny technical teams can become similar.
One person knows a subsystem cold. Another has better product instincts around a certain surface. Another is the person you want around when a machine disappears or a performance number smells wrong. Everyone still knows enough of the product to understand why those specialties matter.
Research on transactive memory gives a more formal version of this. In a classic experiment, groups trained together performed better on a shared assembly task than groups whose members trained separately, and the advantage was mediated primarily by the group's knowledge of who knew what.[^5] Meta-analytic work on shared mental models also finds positive relationships with team process and performance.[^6]
In plain English: part of becoming a team is learning the other brains.
After enough shared work, one sentence can contain a ridiculous amount of history.
“This smells like the stale-state thing.”
“The UI is telling me about the container again.”
“Use the boring path first.”
“Same failure mode as Tuesday.”
A stranger needs the explanation. A close working group receives the compressed packet.
This compression is one of the hidden assets of a stable small team. Replacing people casually throws some of it away. Growing too quickly dilutes it because new pairwise relationships appear faster than shared history can accumulate.
Service and prep are different clocks
The kitchen analogy gets dangerous if “high performance” turns into permanent emergency.
A restaurant has service, and it has prep.
Service is immediate execution. The orders exist. Timing is live. Everybody communicates directly. The standard has to survive pressure.
Prep is where knives get sharpened, ingredients get portioned, equipment gets repaired, experiments happen, and tomorrow's service becomes possible.
A tiny startup needs both clocks.
There are periods where the correct behavior is brutally direct: ship, fix, answer the user, restore the machine, close the loop. There are also periods where somebody should spend half a day asking why the company keeps doing the same stupid thing at all.
Permanent service mode creates a team that gets very good at sprinting in the wrong direction. Permanent prep produces elegant theories with nobody waiting for the result.
The useful person can move between them. They can deliver under pressure, then come back later and ask which repeated pain deserves to become software, policy, a test, a better interface, or simply a different habit.
High standards work better when bad news can move quickly
Intensity creates another requirement: people need to be able to surface ugly information early.
The psychological-safety literature is useful here because the term often gets flattened into “everyone should feel comfortable.” The original work is more operational. Psychological safety was defined as a shared belief that interpersonal risk-taking is safe, and it was associated with learning behavior in teams.[^7]
A tiny company needs the ability to say:
I broke it.
I think we are wrong.
I do not understand this.
I think the last two weeks went into the wrong problem.
I can make this work, but I cannot currently prove it is safe.
A group with high standards and low information flow becomes brittle. People hide mistakes, delay uncertainty, and defend old work because admitting the problem feels more expensive than carrying it.
A group with high trust and a weak quality bar becomes pleasant and mediocre.
The attractive combination is harsher and kinder at the same time: serious standards, quick correction, and very little social tax on telling the truth.
AI looks like an invisible brigade
Now add machines.
A five-person company used to run into a physical labor ceiling pretty quickly. Even extraordinary people only have so many hours. A surprising amount of company growth historically meant adding human nodes because the work itself required more hands.
Generative AI has already produced measurable productivity gains in several controlled or large-scale settings. One controlled developer experiment reported a 55.8% faster completion time on a programming task for participants using an AI pair-programming tool.[^8] A preregistered experiment on professional writing tasks found 40% lower completion time and 18% higher output quality with generative AI assistance.[^9] A later large workplace study of more than 5,000 customer-support agents found about a 15% average productivity increase, with larger gains among less experienced workers.[^10]
Those studies establish bounded productivity gains. The five-person billion-dollar company remains an extrapolation. The effects vary, and organizations contain plenty of work that resists automation or becomes newly important once other work gets cheaper.
The evidence is enough to make the direction interesting.
Imagine every chef suddenly gets a large set of tireless prep cooks.
The chef's comparative advantage moves away from chopping every onion personally. More of the scarce work becomes choosing the dish, sequencing the work, checking quality, resolving ambiguity, spotting the bad batch, deciding what deserves another attempt, and knowing when the whole menu is wrong.
A small AI-native company can begin to look like that.
The humans own judgment, taste, responsibility, relationships, high-consequence decisions, and the weird synthesis that appears when several domains collide. Machines absorb more drafts, searches, code generation, routine analysis, testing, transcription, comparison, bookkeeping, and other pieces of execution.
The new bottleneck becomes coordination between human judgment and abundant machine effort.
Which jobs deserve an agent?
What evidence makes an answer trustworthy?
What should stay quiet when everything is healthy?
Which failures need a person immediately?
What should the system remember so the next worker does not rediscover it?
How do five humans keep a coherent picture when a hundred machine processes are producing artifacts around them?
A traditional company often scales by adding more people, then adding managers to coordinate the people, then adding systems to coordinate the managers. An AI-native company has a chance to move some of that scaling burden into software earlier.
The work ledger can remember who owes what. Tests can carry the quality bar. Receipts can preserve evidence. Automation can handle routine state transitions. Search can recover the past. Agents can do parallel investigation. A good workbench can reduce a large machine population into the few decisions that actually require a person.
Human attention remains expensive, so the machinery should earn the right to interrupt it.
Small teams need internal difference
Five copies of the same person would be a terrible little company.
A tiny group needs enough agreement to move and enough difference to see more than one future.
The shared part can be operational:
We tell each other bad news quickly.
We care about the user.
We finish things.
We can argue without turning the argument into a relationship problem.
We measure claims when measurement is available.
We clean up after ourselves.
We take ownership when the problem crosses our desk.
The differences can remain huge. One person can have stronger aesthetic instincts. Another can think in systems and failure recovery. Another can understand distribution or customers better. Another can have a better nose for product scope. Another can be the person who keeps asking whether the company is about to spend six months recreating a commodity.
Same game, different players.
The phrase “culture fit” becomes dangerous when it means familiarity, sameness, or liking the same jokes. Work compatibility is the useful part. Can these people form one high-bandwidth unit without becoming one narrow mind?
Small teams are fragile too
The romantic version needs its unpleasant half.
Five people means very little redundancy. A departure can remove a giant fraction of the company's memory. Illness can take out an entire specialty. A strong personality can dominate the room. A shared wrong premise can infect nearly every decision because there are fewer independent clusters to challenge it. Friendship can make accountability harder. Intensity can slide into exhaustion because everyone sees every emergency.
AI can amplify the same problem. Five humans plus a hundred agents still fail if the hundred agents inherit one bad assumption. More output can mean more confident garbage, more correlated mistakes, and more artifacts than anyone can inspect.
Small therefore needs explicit countermeasures:
- enough overlap that one person's absence does not erase an entire capability;
- durable records for decisions that future people will need to understand;
- independent checks for high-consequence machine output;
- habits that invite disagreement before a premise becomes expensive;
- slack for sickness, recovery, and actual thinking;
- deliberate contact with users and outsiders so the group does not become an elegant sealed jar.
The goal is to preserve the advantages of a little group for as long as the economics allow.
The company can stay socially intimate while becoming economically enormous
Here is the actual speculative leap.
Human social research suggests that close relationships live in small inner layers. Software teams have repeatedly rediscovered practical advantages to small groups. Team research suggests that shared mental models, transactive memory, and interpersonal safety can improve learning and performance. Startup research suggests that early organizational choices can persist. AI research shows that machine assistance can raise individual productivity in real tasks.
Each of those literatures stops well short of “therefore the future corporation has five employees.”
Put them together, though, and a beautiful possibility appears.
Maybe technological leverage lets some companies stop growing at the point where the humans still form one intimate working group.
The people can all know each other.
They can all know the product.
They can all hold meaningful ownership.
Ideas can cross the entire company in one conversation.
Specialties can exist without departments.
The company's memory can live partly in the people and partly in the systems they build around themselves.
Agents can provide the extra hands without becoming extra social layers.
A huge amount of economic output can accumulate around a group that still feels like a band, a lab, a studio, or a very good kitchen.
Success has historically had a nasty habit of destroying the social condition that made the early company fun. Five people become twenty, then a hundred, then meetings appear to coordinate the meetings. The founders who loved making things become managers of an organization whose job is increasingly to manage itself.
AI gives us at least a chance to test another path.
Build the machinery around the little group.
Keep the little group little.
If it works, lunch stays the all-hands.
Sources and useful reading
[^1]: Cognitive resource allocation determines the organization of personal networks reviews evidence for nested personal-network layers around 5, 15, 50, and 150. [^2]: Using obsidian transfer distances to explore social network maintenance in late Pleistocene hunter-gatherers gives a useful table comparing the 5/15/50/150 layers with family, foraging-party, residential-band, and wider aggregation scales. The mapping is approximate; treat it as an analogy with loose boundaries. [^3]: Small is beautiful: A study of packaged software development teams examined 74 product-development teams and reported a median team size of five. [^4]: Building the Iron Cage: Determinants of Managerial Intensity in the Early Years of Organizations studied young technology firms and found enduring effects from founding conditions and early employment models on later managerial intensity. [^5]: Group Versus Individual Training and Group Performance: The Mediating Role of Transactive Memory found better recall and assembly performance in groups trained together, with transactive memory mediating much of the advantage. [^6]: Measuring Shared Team Mental Models: A Meta-Analysis synthesized 23 studies and found shared mental models positively related to team performance across measurement approaches. [^7]: Psychological Safety and Learning Behavior in Work Teams studied 51 work teams and linked psychological safety with learning behavior, which in turn mediated performance. [^8]: Controlled AI pair-programming productivity experiment reported 55.8% faster completion on a bounded JavaScript programming task for the treatment group. [^9]: Experimental evidence on the productivity effects of generative artificial intelligence found a 40% reduction in completion time and an 18% increase in output quality across incentivized professional writing tasks. [^10]: Generative AI at Work studied deployment across more than 5,000 customer-support agents and reported roughly 15% average productivity gains, with larger gains for less experienced workers.