How a team's output scales with its size

A model of coordination in multi-agent systems, whether the agents are human, artificial or a mix. The formula is the worked example from Imas, Koh and Legg, “Organizing Intelligence” (2026). They apply it to firms and the economy, but nothing in it is specific to firms: it describes any team that divides work and has to coordinate. Build a team from coders, writers, coordinators and AI copies, then see how output per member changes as the team grows.

Organizing Intelligence

Alex Imas, Andrew Koh and Shane Legg (Google DeepMind). arXiv 2611.05417, 31 October 2026.

When will advances in AI lead to economic transformation? We develop a model of how intelligence, the ability to achieve goals, matters for organizations, the key unit of economic production. Agents (human or AI) with multidimensional capabilities sort endogenously into competing firms. A firm’s productivity depends on its agents’ capabilities; we distinguish competence, which shifts productivity levels, from coordination, which governs how productivity scales with size. Two differences between AI and humans drive transformation: AI expertise is reproducible while humans’ are in fixed supply, and AIs may not face the same human limits on coordination. We analyze how equilibrium market structure, productivity, employment, and wages evolve as AI capabilities improve. When AI is more capable than the average human, organizations of complementary humans can nonetheless outperform AI. Increases in AI competence can change the composition of workers within firms and decrease firm size. As AI coordination improves, however, this can break the Coasean ceiling on firm size and the market structure undergoes a phase transition toward large, concentrated AI-heavy firms. Changes in human employment are gradual when AI displaces workers within mixed human-AI firms, or sudden when AI firms outcompete human firms. We decompose wage changes into a technology effect, which changes the value of humans within existing firms, and a reorganization effect as firms adjust size and composition. Reorganization can reverse the technology effect on wages.

Abstract quoted from the paper (CC BY 4.0). Open on arXiv · Companion essay
output per member = L × eφ·m/(1+m) × (1+m)ȳ* × 1/(1+m)   with φ = 7/4, m = team size, L = competence, ȳ* = average coordination
Paper's cases
Team mix (ratio)

Coder: skills 9, 1; coordination 0.75. Writer: 1, 9; 0.75. Coordinator: 1, 1; 0.97.

The AI's profile

Your teamHuman teams from the paper
Show the values as a table

Measured on real agents: the job board

Teams of 1 to 16 copies of one model pick and solve items from a shared board with no communication (run 16 November 2026, four groups per point). The fitted exponent is how output grows with team size, which is what ȳ* is in the formula above for a large team. “Use in model” sets an all-AI team to that value.

Backend Board shown to each agent

Job board: the allocation task

It tests whether copies of one model spread themselves over independent work when they cannot talk.

  • Board: 48 items, each a sum of two two-digit numbers with an opaque ID.
  • Teams: 1, 2, 4, 8 or 16 copies of one model, for 3 rounds.
  • A round: every agent picks one open item and submits an answer. All choices are simultaneous and there are no messages.
  • Closing: an item closes when someone answers it correctly. Several agents on one item close it once, so the duplicates are wasted. Closed items leave the board.
  • Memory: none. Each agent in each round is a fresh call that sees only the prompt below.
  • Score: items closed after 3 rounds, divided by 3 × team size. The best possible is 1.
  • Conditions: the player-number line is present or absent; the board is listed in the same order for everyone, or shuffled separately for each agent in each round. The shared objective is always stated. A lone agent gets a one-agent wording.
  • Runs: Sonnet 4.6 and Sonnet 5.5, four groups per point. Two backends, never pooled: the API with no system prompt at medium effort, and claude -p (Claude Code headless, its default system prompt, no tools).
  • Random choosers: each picks uniformly among the open items and always answers correctly (simulated).
  • Exponent: the slope of log items closed on log team size across the five sizes.

Prompt for agent 1 of 16 in round 1, with a player number. The real board lists all 48 items; six are shown.

You are one of 16 agents working on a shared job board. The board lists open items; each item is a small sum. The work runs for 3 round(s). In each round every agent chooses one open item and submits an answer to it. All agents choose at the same time and cannot communicate.

An item is closed when at least one agent submits a correct answer to it. If several agents choose the same item in a round, it is still closed only once.

You are agent 1 of 16. The agents are numbered 1 to 16.

All 16 agents share one objective: to close as many items as possible in total.

This is round 1 of 3. 0 item(s) have been closed so far.

Open items:
ZX3: 60 + 43
PR7: 30 + 82
RC3: 12 + 69
NX5: 21 + 53
JX5: 16 + 80
QD7: 46 + 28

Reply with exactly two lines and nothing else:
ITEM: <item id>
ANSWER: <number>
Harness: job_board.py. Replies are parsed and scored; raw replies are not stored.

Job board: the shared-knowledge task

It tests whether output per agent can rise with team size when knowledge is costly to produce once and free for everyone to reuse, which is the mechanism behind the paper's supercooperative case.

  • Board: 48 ordinary items worth 1 point each, and 32 locked items worth 4 points each, in 4 families of 8.
  • Research: an agent can spend its round researching one family instead of solving an item. A family unlocks for everyone from the round after it has received 3 research contributions in total.
  • Teams: 1, 2, 3, 4, 8 or 16 copies of one model, for 3 rounds. One action per agent per round, simultaneous, no messages, no memory.
  • Why returns can increase: unlocking is a fixed cost paid once and used by all. Best play scores 3 points for a lone agent, 4.5 per agent for two, 8 for three and 8.25 from four up.
  • The coordination problem: the team needs about three researchers on each family, not none and not everyone on one.
  • Conditions and runs: as in the allocation task. With shuffling, the families and the items are listed in a different order for each agent.
  • Random choosers: each picks an available kind of action uniformly (research, solve an ordinary item, solve an unlocked item), then a uniform target (simulated).

Prompt for agent 1 of 16 in round 1, with a player number. The real board lists all 48 ordinary items; five are shown.

You are one of 16 agents working on a shared job board. The work runs for 3 rounds. In each round every agent takes exactly one action. All agents act at the same time and cannot communicate.

There are two kinds of item, each a small sum:
- Ordinary items are worth 1 point each. They can be solved at any time.
- Locked items are worth 4 points each. They belong to families. A locked item can be solved only after its family has been unlocked.

A family is unlocked by research. Researching a family uses your whole action for the round. A family becomes unlocked, for all agents, at the start of the round after it has received 3 research contributions in total. Contributions can come from any agents in any rounds, and they add up across rounds. Contributions beyond 3 are wasted.

An item is closed when at least one agent submits a correct answer to it. If several agents choose the same item in a round, its points are counted once.

You are agent 1 of 16. The agents are numbered 1 to 16.

All 16 agents share one objective: to score as many points as possible in total.

This is round 1 of 3. Points scored so far: 0.

Families:
QR: locked, research 0 of 3 (8 items)
CL: locked, research 0 of 3 (8 items)
VT: locked, research 0 of 3 (8 items)
QM: locked, research 0 of 3 (8 items)

Ordinary open items:
TP5: 75 + 28
MG3: 43 + 79
ZG6: 23 + 20
NT3: 56 + 66
NZ5: 81 + 72

Reply with exactly two lines and nothing else. To solve an item:
ACTION: SOLVE <item id>
ANSWER: <number>
To research a family:
ACTION: RESEARCH <family>
ANSWER: none
Harness: job_board_research.py. Replies are parsed and scored; raw replies are not stored.
Allocation: share of the maximum output

Items are independent, so the best possible is 1 at every size and the exponent cannot pass 1. The blue line is the formula's coordination term at your team's current ȳ*: move the coordination slider to fit a measured line.

Shared knowledge: points per agent

Valuable items unlock for everyone after three agents research their family. A fixed cost shared by all lets output per agent rise with team size, if the team divides the research.

Sonnet 5.5, told its player numberSonnet 5.5, no numberSonnet 4.6, told its numberSonnet 4.6, no numberRandom choosersBest playFormula at your ȳ*

Limits: the allocation task measures only losses, so it shows how close a team gets to the threshold of 1, not whether it can cross it. Reading its exponent as ȳ* is a modelling choice: the paper's friction is attention split among colleagues who talk; here it is collisions among agents who cannot. Four groups per point, two models from one provider, trivial items.

Why there is a best size

Dividing work helps, then levels off. Coordination helps or hurts depending on whether ȳ* is above or below 1. Their product is the scale effect.

Dividing workCoordination, net of split attentionProduct

Why mixing specialists raises competence

Competence L is the harmonic mean of the team's average coding and writing skill, so the weaker skill holds it back. Curves join skill pairs with equal L.

What is and isn't from the paper

The formula, the three human types, φ = 7/4, the AI rental cost of 1 per copy and the deep-market price of 0.3 are the paper's (Section 2). The authors call this functional form “a little ‘ad hoc’”; their general model replaces it with a scaling exponent γ. The human cases here reproduce their numbers: 5.0 per coder at size 6, 13.8 for a coder–writer team, 13.0 at size 7 with a coordinator. Where the paper says “firm” and “worker”, this page says “team” and “member”.

The paper calls its unit a firm and uses the model to study market structure and wages. The model itself has no ownership, contracts or liability in it, only members with skills, a size and a coordination cost. So this page reads it as a model of teams: a software team, a swarm of AI agents, or people and agents working together.

The page computes one team at a time. It does not solve the paper's market equilibrium, so it shows revenue per member, not how pay is split between types. The paper's AI-regime figures fix a coordination of 1.60 for the supercooperative case; the task skill of 5 used for that preset is my choice.

The job-board section is not from the paper. It is Steve Phelps's experiment of 16 November 2026 (harnesses job_board.py and job_board_research.py); the exponents are fitted on five team sizes up to 16.

Source for the model: arXiv 2611.05417.