The org chart is upstream of almost every shipping problem an engineering organization will hit. The handoffs that slow releases, the teams that cannot get capacity, the platforms that should have been built two years ago and were not — the root cause is rarely the people. It is the structure they are working inside. Three patterns dominate modern engineering orgs, each with a clear shape, a clear failure mode, and a clear set of conditions where it actually works.
Hub-and-Spoke
A central engineering platform team owns infrastructure, tooling, and shared services. Product teams sit as spokes, consuming the platform and shipping their own surface area. The model works in organizations where platform leverage is the primary unlock and where product surface area is relatively bounded. It fails when the platform team becomes a bottleneck or when product teams route around it because the platform does not match their actual needs.
Team Topologies' team patterns codify the variant that holds up: a thin, opinionated platform team that ships internal product, with explicit service-level commitments to its consumers. The version that fails is the one where the platform team is graded on elegance rather than adoption.
Embedded
Embedded models put specialists inside product teams — an SRE in this team, a security engineer in that one, a data engineer in another. The model wins when the work requires tight collaboration with a specific product surface and where central queues create unacceptable latency. It fails when the embedded specialists drift from the central practice, lose connection to peers, and quietly become generalists.
The fix that works in practice is what Spotify popularized and others copied: keep the embedded specialists reporting to a central craft lead, even as they sit and work inside product teams. The dotted line is doing more work than it looks like it is.
Federated
Federated models scale by giving product teams full autonomy over their stack, with a small central team owning standards, contracts, and cross-cutting platforms. The model works for organizations whose product surface is wide enough that no central team could keep up with the variance. It fails into chaos when the central team is too thin and into central planning when it is too thick. The calibration is harder than it sounds.
The DORA State of DevOps research consistently finds that federated organizations outperform centralized ones on delivery velocity — but only when the central platform layer is genuinely good. Federated autonomy on top of a weak platform produces faster duplication of bad decisions.
How To Choose
The choice is rarely categorical. Most organizations beyond a hundred engineers run a hybrid: a hub-and-spoke platform for the load-bearing infrastructure, embedded specialists for the cross-cutting disciplines, and federated product teams shipping on top. The questions worth answering before the org chart is drawn: what is the platform unlock worth? Where is the latency cost of central queues unacceptable? And what is the central craft strong enough to standardize? Answer those, and the structure follows.
Common Structural Mistakes
Three mistakes show up across organizations that stall. Standing up a central platform team and grading it on elegance instead of adoption. Embedding specialists without the dotted line back to a craft lead, and watching the specialty erode. And declaring federation before the platform layer is solid, which produces faster chaos rather than faster shipping. Each of these is recoverable. They just rarely get recovered until somebody names the failure.
Key Takeaways
- Hub-and-spoke wins when platform leverage is the primary unlock and product surface is bounded
- Embedded wins when central queues create unacceptable latency — keep the dotted line to craft leads
- Federated wins when product surface is wide and the platform layer is genuinely good
- Most orgs at scale run a hybrid: platform for infra, embedded for craft, federated product teams on top
- Grade platform teams on adoption, not elegance
- Federate only after the platform is solid — federated chaos is just faster wrong