When 1 + 1 > 2

2026-01-31


From my time at Greptile, I witnessed a team that achieved an astonishing amount. By that I mean the mixture of different members produced far more than what would have been possible with a team of the same size composed of clones of any single member. Unfortunately, I have also observed the contrary: teams that produce a questionable amount of output compared to what each member seems capable of alone.


Having had to run engineering teams, I needed to understand why this happens. I couldn’t help but feel that the existing literature on the topic wasn’t systematic enough: there is a lot of great tactical advice, but no clear theory about the underlying nature of the phenomenon. Based on what I have observed and read, I want to propose a simple model of team output in terms of skills, attention allocation, and the dynamics of a multi-agent system.


Tasks, Skill, Productivity, and Output

Before we measure the output of a team, we start by defining the output of an individual. Broadly, an output is a change in the state of the world between two points in time. A task is the desire or plan, by some system, to move the current state of the world toward a desired goal state. Here, the system we are dealing with is the individual: an individual can work to produce output. A non-constructivist observation I’ll make about tasks is that they can be decomposed into subtasks that reveal their structure and, as we will show below, the concrete skills necessary to tackle them.


The concept of skill level is simple: it is some affinity metric attributed to an individual with respect to a task. Skill is probably a product of talent and experience; however, for this model it isn’t too important to define exactly how one achieves a given level of skill. I am merely claiming that there is some comparative measure between people in a subject area. I don’t think this is controversial.


Bringing the concepts of task and skill together, we can think of a task as a bundle of skill-demands. And an individual’s skill level with respect to a task is inversely related to the time required to complete that task: the more skilled you are at a task, the faster you will get it done. Some tasks are more complex than others; for such tasks, it is helpful to decompose them into simpler subtasks. Doing so helps us not only understand the task better but also understand the skills required to complete it effectively. For example, the skill of building a software company is broad. Breaking it down into engineering ability, product intuition, the ability to sell, and so on helps us understand both the nature of building and selling software and the types of individuals necessary to do it efficiently.


With productivity I want to capture two properties: that it is finite and that it varies. Defining productivity as the allocation of one’s attention, scaled by an efficiency parameter, captures both. Attention is a finite resource bounded by time. The efficiency term captures factors such as flow states, degree of sleep deprivation, and so on. This efficiency fluctuates over time and varies between people. Thus, productivity is essentially how much of your skill “projects onto” a task per unit of attention-time. If a task requires programming and you’re a great programmer, your productivity on that task is high. If it requires sales and you’re terrible at sales, you won’t make much progress. A familiar observation from everyday life is that skill level increases over time with conscious practice (in other words, attention allocation). Attention, skill, and task are all you need for change.


Thus, we can redefine output as productivity accumulated over time. This maps nicely to the definition of work in mechanics: output is the integral of productivity, like power, over time. A task is “complete” when cumulative output equals the total displacement required by the task. This model captures the intuition that people who are more skilled and more efficient generate more output in the same unit of time.


Organizational Dynamics

The model of output for an individual is straightforward; now we try to model what happens when individuals start working together. The primary intent of modeling team output, as alluded to in the introduction, is that in the real world we observe teams getting more, the same, or less work done compared to the “cloned productivity” of any member. This suggests that there are hidden dynamics between members that affect the final team output.


We can naively introduce an efficiency parameter that acts as a positive or negative multiple on the combined productivity of each member. However, this isn’t very intuitive, and it doesn’t really capture how such a parameter behaves to produce dramatically different results or why it exists in the first place. I believe the most important variables in team dynamics are how a team allocates each member’s attention and how it mitigates the penalties of decentralization.


Member Attention Allocation

The key insight is how attention reallocation changes individual productivity. When working alone on a task, you must divide attention across all required subtasks even those where your skill is low. Your effective productivity is a blended rate, dragged down by any gaps in the skills required for the overall task. For example, if the task of building a software company is 49% programming and 51% sales, and you’re great at programming but terrible at sales, that 51% will hurt your productivity.


In a team whose members have complementary skills, attention can be reallocated more effectively. Each member focuses on work where their skill is high, and tasks where their skill is low are handled by those for whom it is a strength. Perfect specialization is perfect attention allocation: team productivity approaches a state where every subtask is handled by whoever is best at it. In that idealized setting, this is greater than or equal to the solo case for any non-trivial team. This is where the superlinearity in team productivity comes from. Two people working alone each produce at their blended rate. Two people working together with complementary skills can each produce closer to their peak rate. In this sense, 1 + 1 > 2 because “1” alone is not spending all their time doing what they are best at.


This relationship is well documented in founding teams of startups. The inventor + entrepreneur combination is an established archetype in startups because of this reason. When you have a task that requires skills rarely found in a single individual (though exceptions do exist), you want two people who are world-class in each area to complement each other in getting something started.


This sounds great, and the sensible next step is to identify the skills needed for a task and hire as many people as possible to maximize the total attention-skill available. However, anyone with operating experience knows this isn’t true in general. You can have a large group of extremely smart people and still fail to produce the output of a much smaller team, a quirk in the system that explains why small companies and startups stand a chance against big companies at all. The nature of this phenomenon, I suspect, is tied to the difficulty of working with other people.


Coordination Overhead

Working with others is hard. To successfully reap the benefits of a highly skilled team, the team must coordinate properly to prevent deterioration in productivity.


The theoretically optimal attention allocation assumes perfect information and perfect allocation: each subtask goes to whoever is best at it. To know that, at least one agent must determine the optimal allocation and get others to agree to it. Clones of oneself could achieve optimal coordination trivially; since each clone has identical skills, the question of “who’s best” is irrelevant. But working with others requires first discovering who is actually best at each skill (or is expected to grow in that skill) and then implementing that allocation. If this process breaks down, you get suboptimal assignments. Even worse, you may introduce gaps in attention allocation when a necessary task is unowned by any member.


Members within an organization can have, and often do have, misaligned incentives. Someone optimizing for their own promotion or their faction’s influence is not fully contributing to the team’s objective. Misaligned incentives don’t just reduce team productivity through bad allocation; they can work against it. A manager who claims ownership of a task to grow their team’s scope pushes allocation toward political power rather than toward necessary work. They can grab high-visibility tasks regardless of fit and leave less glamorous but necessary work unowned or worse, prevent anyone else from owning it. Teams can refuse to let others work in “their” area, and can even deliberately sabotage. In the limit, this corrupts not just allocation but the desired output itself. It is also important to note that a person can be highly skilled, low-maintenance, and still net-negative through misalignment. Misaligned incentives make team dynamics and output modeling nearly impossible, and they aptly act as an agent of chaos.


An underrated overhead in coordinating teams is communication. Communication is a skill in itself, and therefore requires attention. This overhead scales with the number of communication channels each member must maintain. A commonly presented model assumes coordination cost grows with pairwise communication links equating to roughly O(n2)O(n^2) in dense groups. In practice, the communication overhead most often includes other sub-teams as well as the context of the teams does not necessarily equate to each the context of each member. This cross-team dependency grows the coordination surface area combinatorially. In a group of nn members there are O(2n)O(2^n) possible sub-teams where now it could mean that you end up with O(22n)O(2^{2n}) possible communication channels. The communication problem is a lot larger than what people give it credit for.


How can we mitigate this? The general strategy roughly seems to be in understanding the team and the task, structuring the organization, and setting shared culture. Understanding the team and the task is the first step in properly allocating attention; computing a good allocation requires knowing everyone’s skill profile and knowing what skills the task demands. Good structure simplifies the coordination problem. With proper partitioning of tasks and teams, it becomes clearer who should be allocated where, and the number of required communication channels reduces to collaborators and the manager you report to, essentially sparsifying the communication graph. Finally, a shared culture reduces the need for communication and coordination altogether. If people are on the same page, the overhead per channel drops and incentives align naturally. Shared context means less time spent explaining, trust means less time spent verifying, and common goals mean less attention lost to unproductive causes. This seems to be inline with the literature in what good managers are supposed to do.


Organizational Output

Putting it all together: team productivity for a task is determined by how attention is allocated amongst members of varying skill, minus the overhead of coordination and communication. The output of a team is then the same as the output of an individual: it is the integral of team productivity over time. On this view, organizational design is an optimization problem: maximize the gains of specialization while minimizing overhead cost of coordination, alignment and communication. The practical job is to structure teams, processes, and interfaces so the coordination tax stays below the specialization surplus, thereby minimizing time-to-output.