communicoupling
Any organization that designs a system will produce a design that copies the organization's communication structure. Mel Conway wrote that in 1968, and his own example still stings: assign a compiler to four teams and you will get a four-pass compiler. The mechanism is plainer than the slogan. Every design decision that spans two parts needs coordination; coordination travels the channels the org already has; where a channel is thin, the two parts stop negotiating and settle for a hard, minimal contract. The interface is the fossil of the conversation nobody could afford.
Below, twenty-four modules with fixed functional interdependence are dealt to a number of teams you choose. Each round, coupled pairs need a coordinated change: inside a team it lands across a desk; across teams it costs meetings the teams may lack the bandwidth to hold, and every coordination that fails ratchets that pair's interface toward a frozen contract. The Conway ratio — cross-team hardness over same-team hardness — reads how legibly the org chart is being written into the artifact. Re-deal the teams mid-run and watch the artifact follow.
joint custody · held with MatheMetaphor's homomorphism — the mirroring is a structure-preserving map: who-talks-to-whom in the org carries over, seam by seam, into what-calls-what in the artifact. That page has the mathematics of maps that keep structure; this one has the machine that does the copying.
the mechanism
The model keeps the problem fixed and moves only the org. A random coupling matrix says which of the 276 module pairs genuinely need co-change; each round a pair needs a coordinated change with probability proportional to its coupling. Same-team coordination costs 1 and always lands. Cross-team coordination costs 1+f, split between the two teams — and each team's meeting bandwidth scales with its headcount, so a three-person contractor shop can host far less outside talk than a room of twelve. When either side's budget is spent, the coordination fails, and the pair's interface ratchets harder by +0.22·(1−H); when the talk happens, it relaxes by −0.12·(H−0.15). That single asymmetry writes the org into the artifact. With three squads at friction 2, same-team seams settle at hardness 0.18, cross-team seams at 0.63 — a Conway ratio near 3.4, from dynamics that never once look at the org chart directly.
Hardening feeds itself: a hard contract absorbs a change without a meeting, so demand for cross-team talk falls as the seam freezes — in the model, a fully hard interface fields 60% fewer coordination requests. That is why the mirror, once formed, is quiet. And it is why the inverse maneuver works. Re-deal the same modules into new teams mid-run and the ratio, read against the new org, collapses to 1.01 at the stroke — the inherited architecture matches nobody's boundaries. Then the machine resumes copying: the new ratio rebuilds 80% of its old value within about 350 rounds, while the ghost ratio — hardness read against the previous org — fades from 3.0 to 1.08 within 300. If you want a different architecture, the cheapest editor is the org chart.
what to try
Keep three squads and sweep friction from 0 to 5 (restart between runs). The Conway ratio climbs from 1.2 through 2.4 and 3.4 to 4.2, as cross-team hardness rises from 0.22 to 0.79. Same teams, same problem — the artifact records only what the talking cost.
Let squads harden, then hit Reorg teams — or drag the team dial, which reorganizes live. The gold line crashes to ~1.0 and grinds back up while the blue ghost of the old org decays toward 1.0. Conway's law, run in reverse: the architecture follows the org, about 350 rounds behind it.
Compare spend across presets: distributed contractors pay about 21 per round to the squads' 29. The worst org is the cheapest, because 71% of its needed coordinations never happen — they are "resolved" by hardening instead. Low meeting load can mean health, or it can mean the system has given up talking.
the law, everywhere
The evidence arrived decades after the aphorism. Herbsleb and Grinter followed a product split between sites and watched the corridor channel die: the casual question that costs nothing across a hallway becomes a scheduled call across an ocean, so changes that crossed the site boundary crawled, and the seams between sites hardened into formal handoffs exactly as the model's ratchet predicts. The mirroring hypothesis studies made it quantitative: co-change that spans a module boundary costs more when it also spans an organizational one, and products built by distributed volunteers come out measurably more modular than the same kind of product built in one building — the loose org could not afford anything but clean seams, so clean seams are what it made.
Once you believe the mechanism, the design order reverses. The org chart stops being an administrative afterthought and becomes the first draft of the architecture — so teams get drawn around the seams you want, before the seams exist. That is the inverse Conway maneuver, and it is now orthodoxy in software: decide the boundaries the problem wants, then staff them, then let the ratchet do the rest. The model shows both its power and its bill. Architecture does follow the org — but hundreds of rounds behind, through a valley where the inherited seams match nobody's boundaries and the ghost of the old structure still collects coordination taxes. Reorganizing is rewriting the artifact with the slowest editor there is; the law only guarantees it works eventually.
the mapping
| In the model | In the world |
|---|---|
| 24 modules | The parts of the artifact — services, chapters, wings of the building, clauses of the treaty. |
| the coupling matrix | The problem's own interdependence: which parts genuinely need to change together, org chart or no. |
| team assignment | The org chart — this quarter's. The one structure the designer actually controls. |
| friction f | Distance in every currency: time zones, contract boundaries, security clearances, mistrust, process. |
| per-head bandwidth | How much outside coordination a team can physically host before requests start bouncing. |
| hardness H | What the seam is made of: shared understanding at the soft end, a frozen minimal API at the hard one. |
| the conway ratio | The mirroring itself — how legibly the communication structure is written into the design. |
where it tears
Here the org shapes the artifact and never the reverse. In life, teams congeal around existing components, hiring follows the architecture, and the two co-evolve until cause and effect are archaeology. The model also freezes the coupling matrix — real functional interdependence drifts, and yesterday's correct seam becomes today's fault line without anyone reorganizing anything. Conway's law is a strong tendency in a loop, and this page simulates only its forward half.
Conway said communication structure, and the model betrays him slightly: its org is a clean partition plus one friction scalar. Real coordination flows through back-channels, lunch tables, ex-teammates, and the one engineer who knows everyone — brokers whose private channels soften seams the formal chart says should harden. Where the informal graph is rich, the mirror blurs; measure the org's actual talk, not its chart, before predicting its architecture.
The instrument's reds and the word "failure" make hardness read as damage. Half the time it is engineering: hard minimal contracts are how large systems stay buildable at all — near-decomposability, purchased on purpose. The pathology is misplacement, a hard seam laid across a strongly coupled pair that needed to keep talking. The ratio measures how faithfully the org is mirrored, and fidelity to a bad org chart is precisely the disease.