Keith MacKay

Posted on Jun 7 • Originally published at tlcmentor.substack.com

The Mythical Management Month

#agents #ai #leadership #management

15 Direct Reports and The Mythical Man(agement) Month

Everyone is redesigning their org for AI...but the ones who get it right will remember to design for the humans involved.

Brian Armstrong just announced Coinbase is restructuring around small, nimble teams and managers with up to 15 direct reports [1]. The logic is appealing: smaller teams move faster, AI extends individual output, fewer layers means clearer accountability. I don't disagree with any of that.

But 15 direct reports, 30-minute 1:1s, once a week: that's 7.5 hours. A whole day. And it won't pencil out that cleanly -- other meetings and commitments will get in the way and things will get moved, cancelled, or become interruptions of important blocks of productive working time.

Paul Graham wrote an essay in 2009 about the fundamental challenge these Coinbase managers will face: the maker's schedule and the manager's schedule are fundamentally incompatible [2]. Makers (knowledge workers) need long uninterrupted blocks for productive work; a single 30-minute meeting doesn't cost 30 minutes, it costs the half-day on either side of it that is lost: losing one's place to jump to the meeting, then rebuilding that context from the ground up after the meeting. [As an aside: we see the same thing when we clear the context window in a coding agent...our team uses tools to store and restore session context before and after the clear to reduce the pain. I have not found a way to do that for human context effectively...and a 30-minute meeting will reliably clear my ADHD context window.] Coinbase's model asks these managers to be player-coaches who both lead and build -- which means running a manager's calendar while trying to hold a maker's schedule inside it. That's not a workflow optimization problem. It's an unsolvable math problem (Graham solved it by running two shifts: manager's schedule before dinner, maker's schedule after dinner...not unusual for entrepreneurs, but unsustainable for many people).

Move the 1:1s to biweekly and you've "only" spent half a day per week, but now you're seeing each person for 30 minutes every two weeks. Monthly? You've clawed back most of your calendar, but you're investing only 30 minutes per working month in the people you're supposed to be developing. At some cadence between "constant" and "never," there's a schedule that sounds efficient and quietly breaks everything.

This is the org design problem that is getting left out of the "gotta use AI" frenzy: connection matters -- and it's a cost of doing business that you can't always engineer away.

Brooks Had a Point

Frederick Brooks published The Mythical Man-Month in 1975 [3]. To oversimplify his central observation: adding engineers to a late project makes it later. Some work is inherently sequential. The example he uses in the book is that a baby takes nine months regardless of how many people you assign to the task.

Management has sequential, non-parallelizable components too. You can use AI to prep for your 1:1s faster. You can use note-capture tools and second brains with agentic AI assistance to log the things you'd otherwise forget (and I highly recommend it!). You can summarize performance data, flag patterns, auto-draft your weekly update. That's real. Those savings add up to hours.

What you cannot parallelize is the relationship itself.

A manager who reads their AI-generated brief right before a 1:1 -- "Taylor had a tough sprint, prefers directness, is working toward a promotion conversation in Q3" -- knows Taylor...but only in the same way that reading an IMDB page means you know an actor. Interesting information. Useful context. Not the same as the thing it's describing.

Dunbar's Number helps to shed some light here. Robin Dunbar's research on primate cognition and human social groups found that we maintain meaningful relationships in layers: roughly 5 in our innermost circle, 15 in the next, 50 beyond that [4]. These aren't preferences. They're cognitive load limits. Dunbar himself said the innermost 5 people (spouse, immediate family, best friends, closest work colleagues) receive roughly 40% of your total social time and emotional capital [4]. Relationships at the 15-person layer are real but thinner; they require consistent investment to stay functional.

Note the coincidence: the "up to 15 direct reports" number sits right at the outer edge of the layer where humans naturally maintain close working relationships. Push past it and you're not just adding meetings. You're asking people to care meaningfully about more people than their brains are built to track.

I've seen this in teams, and I've noted before that Dunbar's number correlates quite nicely to when we need layers of management in a start-up to continue growing:

With 5 people: you're family. Everybody is self-managed and knows what everyone else is doing, often minute-to-minute, without formal meetings. You have shorthand with each other so "meetings" are two-minute knowledge transfers over a donut.
With 15 people, you're extended family. You have moved to multiple people in some roles, and you need to start having managers. You still know everyone, and what everyone is doing, but meetings start happening to allow distribution of tasks, organization of the work. You know your cousins, but you don't know all of them well. How would you?You spend much less time with each of them.
When you get to 50 people, it's more like the family reunion. You recognize them all. You can't know what everyone is working on today and do your own job well. Meetings (and everything else) need to be more formal, so they can be managed and so strategy is transferred from top management down the org chart (now there are tactical layers that aren't management). There are many more middle managers.
At the next level, the 125-150 level, we start to see the largest size group where humans seem to be able to manage ongoing relationships at some level. Larger than this, we just need additional layers of management to aggregate the reporting up and the messaging down. One aspect of Dunbar's work showed that when human tribes grew to 125-150 members, they would split into multiple tribes. It seems to say something about the human brain's absolute capacity to manage and maintain relationships with other humans. After all, we evolved in small tribes with limited mobility and shared value in building relationships to work together for survival. We're literally built for that.

So much value arises from good relationships. Having each other's backs -- not just loyalty, but understanding of what teammates need (which is NOT the same from day to day...to the chagrin of new managers everywhere). Cutting each other slack on a bad day. Understanding each other well enough to SEE that it's a bad day. Real trust in each other -- trusting teammates with things that are going on at home that might get in the way of work on occasion is a tricky business. Many will say work is NOT the place for that (and I'm not suggesting you pull a colleague aside right now and tell them about home struggles). And yet, we all have lives outside of work, and bad days...the very best jobs I've had are the ones where small, high-performance teams knew each other well enough to be ABLE to share things about their own lives and struggles with each other where it would have a work impact -- and the team could plan and function better as a result. Clear communication with respect, love, and candor for each other is the goal. This requires trust, which only comes from time together.

Good relationships are an investment that requires time and consistency.

The Number Isn't Fixed. The Range Is.

Before the "but Dunbar's Number is contested!" crowd starts in: yes, the exact figures are debated. Swedish researchers re-running Dunbar's analysis with updated statistical methods found confidence intervals so wide that specifying any single number is, in their words, "of limited value" [5]. Individual variation is real and documented. Extroverts naturally maintain larger networks than introverts. I would argue the small business numbers for management layers are around 10, 25, 50, 100, perhaps because work relationships are a bit shallower than non-work relationships. The outer relationship layer (the ~150 figure) appears more stable cross-culturally than the inner ones.

But any variation is a range, not an escape hatch. A 2024 study examining how people allocate emotional energy across relationship layers found that some individuals commit 45% of their social attention to their inner 5, others only 15%. Either way, NOBODY is managing 30 deep relationships simultaneously [6]. The cognitive load may not be uniform, but it IS bounded -- we each have limited time and energy available, and must choose how to spend those limited resources on our relationships.

Will an AI-native generation learn to stretch these limits? If you've grown up coordinating dozens of social connections through apps, does your brain actually develop a wider working memory for relationships? We don't know yet, but I suspect we'll find that there ARE brain changes from our new learning and exposure patterns (we've seen this in existing younger generations who have grown up with social media, for instance). We DO know that the research on depth vs. breadth consistently points one direction so far: organizations where managers can build high-quality relationships with their reports show meaningfully better trust, knowledge sharing, and team performance than those running wider spans [7]. A 2025 study found that expanded span of control specifically reduces leadership effectiveness by cutting into the relationship-building time that makes management work in the first place [7]. If your experience as a human being in the world is anything like mine, none of this will surprise you.

The takeaway is that the number of connections varies from individual to individual, but the pattern we've evolved for human interaction doesn't. Trading depth for breadth has costs, and they compound. What we haven't yet studied is what a generation of workers who have only ever known high-breadth, lower-depth management looks like at scale. I think we're going to see the experiment play out around us in real time, and I expect the pendulum will swing back.

The IMDB Problem

I use a second brain. Obsidian, notes on everything, agentic AI that helps to provide context just when I need it. I'd lose track of far more without it. In a 1:1, being able to pull up "we talked about this in December and here's what we agreed" has real value.

But when I'm reading notes to remember something that I feel like I should organically remember in the interest of a relationship, I'm doing something different than managing. I'm performing management. The file has the data. The relationship has atrophied. Do I KNOW this person and what they need, or do I know ABOUT them, the way I know ABOUT an actor from their IMDB profile? Those are different things.

AI can help with the measurable inputs: feedback documentation, promotion narratives, goal tracking, prep for difficult conversations. These are real leverage points and you should use them. What AI cannot do is replace the accumulated texture of time spent. The trust that builds from seeing someone struggle and staying in it with them. The instinct that develops when you've watched someone work long enough to know the difference between "they're quiet because they're focused" and "they're quiet because something is wrong."

The parts of developing people that can be made more efficient with AI are the administrative parts. The actual development happens in the accumulated moments that look inefficient from the outside.

Observations In a Group Chat

Talking about the Coinbase decision in a chat composed of VERY experienced C-Suite veterans, I weighed in that 7 or 8 direct reports is, in my mind, a healthy upper target. I ran mastermind groups in the past, and found that this was also the sweet spot for these groups. Larger than that, and the people in the group didn't spend enough time sharing in the sessions to build the trusting relationships necessary for the room to be a safe space for some truly sensitive and supportive conversations. A phenomenal manager and builder of multiple successful businesses indicated 10 people + 5 agents would be an upper limit for them. The largest number anyone in the thread suggested for a conceivable upper limit was 12, and that came with a qualifier: only works for a superstar (and I might add: "and even then only if the 12 reports are all high performers who need limited direction").

High performers still have bad days. They or their kids get sick. They go through divorces and layoffs and parents in hospice and water leaks and cable installation and the whole catalog of things that happen when you hire humans instead of machines. Being a "superstar" doesn't make someone a robot.

I conduct some new manager training for our group -- helping them understand the differences as they move from managing tasks and work streams to managing teams of people. I tell them what I believe: this is the hardest transition in a career. Not because the skills are technically complex. Because the whole frame has to flip.

Tasks are bounded. You can learn them, master them, build intuition for them, delegate them cleanly. If a task goes wrong you diagnose the problem, fix it, move on.

People aren't bounded. A person having a rough month doesn't come with error logs or reboot instructions, or a right way to solve the problem. The "fix" might take years and involve influences you can't see or control. And you as a manager are also having good days and bad ones, also bringing your own history and blind spots and capacity limits to every interaction.

We're not automating our way out of that. We shouldn't want to.

Teams of Humans and Agents: Who Adapts to Whom?

The conversation is shifting from "AI tools that help individuals" to "teams composed of humans and agents" (or even "dark factory" teams with no humans at all ["zero-person companies"], which is a whole 'nother topic to think through and a fascinating opportunity for some types of businesses). That's real and worth taking seriously.

Research on human-AI teaming points to something fascinating: unlike human teams, where you largely recruit members with relatively fixed capabilities (and only sometimes hire to fill psychological or workstyle gaps/complementarities), AI teammates can be instantiated to match the profile you need [8]. The agent can be configured to prefer directness or to check in more often or to front-load its uncertainty. The flexibility is genuinely remarkable (I predict that we will see the personalities and work styles of agents evolve to complement the strengths and work styles of the particular human team-members with whom they are working).

The research also demonstrates the current limits. Humans have evolved as social animals such that teams negotiate roles naturally and interpret implicit social cues without thinking about it. AI agents need explicit protocols to do the same work. The agent will do what you configured. It won't notice what you didn't configure.

The right frame isn't "will AI adapt to humans or will humans adapt to AI?" The right frame is: what does this specific team need, and who is responsible for maintaining it?

In a well-designed human-agent team, the agent handles the parallelizable, repeatable, stateful work: tracking progress, surfacing context, flagging when something deviates from the plan. The human handles the work that requires judgment, relationship, and the kind of presence you can't stub in a config file.

All-agent teams are coming for certain categories of work. For others -- the work that turns on trust, creativity, and the weird non-rational things humans do under pressure -- the agent is likely the foil and thought partner and collaborator, not the lead.

The Future Is Already Here--It's Just Not Evenly Distributed

This section heading is a favorite quote from William Gibson, popularized over 30 years ago -- and more true now, if anything. Companies are beginning to experiment with the limits of agentic technology at scale (Meta is recording employee keystrokes, mouse movements, and screenshots as a training set for agentic AI that could in theory eventually replace those employees [9]). Will this work? I bet unequivocally yes -- for some cases, with some spectacular failures to come.

Cursor runs $500M in annual revenue with 50 engineers [10]. That's a staggering number, and it's not magic: it's a small team with high context, deep ownership, and tools that extend what each person can do. The math works because everyone carries the whole picture.

But Cursor's model doesn't scale by adding 15 people per manager and calling it nimble. It scales by maintaining the conditions that made small teams effective in the first place: real relationships, shared context, people who know each other well enough to move without constant coordination overhead.

The tools to run a fundamentally different kind of organization exist today. The constraint is human and organizational absorption speed. Testing takes time. Approval processes take time (heavens to Mergatroid do they take time!). Training takes time. The humans in the loop -- not because they're inefficient, but because they're human -- need the time it actually takes to build knowledge, habits, trust in new systems, new teammates, new ways of working. That's not friction to be eliminated. That's the pace of durable change.

You can accelerate the administrative layers. You can learn faster than most do by using better teaching/training/learning techniques. You cannot compress the relationship layers. Organizations that understand the difference will build structures that last. Organizations that don't will run excellent pilots and wonder why nothing sticks.

The Bottom Line

Fifteen direct reports doesn't fail because the math is wrong. It fails because connection doesn't scale linearly. Dunbar figured that out from looking at human communication patterns. Brooks figured it out from observing software projects. Every experienced manager has figured it out the hard way (usually around year two).

The org designs that win in the next five years will be the ones that use AI to ruthlessly eliminate everything that shouldn't require a human, and then protect with equal ruthlessness the time for everything that does. The monthly 1:1 that feels like efficiency now will produce...monthly relationships. And monthly relationships, I would argue, are very different from the relationships you want your teams to have. They're acquaintanceships. You can't build someone's career on an acquaintanceship.

What's your current span of control, and where do you feel the edges? I'm curious whether the 7-9 ceiling holds across industries or whether there are domains where it genuinely breaks.

References

If this resonated, here are some related articles:

For why the layoff wave being attributed to AI is more complicated than the headlines suggest: Are Companies Really Doing Layoffs "For AI"? | Substack
For how to think about managing AI itself — the same relationship-vs-tool tension applies when the "direct report" is an agent: Situational Leadership for AI: More Like a Capable Colleague than a Fancy Formula | Substack
For the org design question underneath this one — how much human judgment to keep in the loop as AI takes on more work: An Evolving Strategy for Knowledge Work: From Human-In-the-Loop to Human-Before-the-Loop | Substack
For why enterprise AI confidence tends to outrun enterprise AI competence — relevant to anyone betting org design on AI capabilities that aren't fully proven yet: The Dunning-Kruger Effect, Now Available at Enterprise Scale | Dev.to | Medium | Substack

_Keith MacKay is a technology strategy consultant and CTO in EY-Parthenon's Software Strategy Group (SSG), specializing in AI disruption and technology diligence for private equity and corporate clients. SSG's AI Disruption Lab conducts rapid assessments of how AI transforms and threatens existing business models and value chains. Keith teaches at Northeastern University and writes about strategy, management, and AI/technology, with Claude Code and Codex as AI collaborators.

DEV Community