A startup's early speed comes substantially from proximity: the people deciding are the people closest to the customer, the code, the problem. Decisions are fast, informed, and contextual. Then it scales, decision-making drifts upward and away from the front line, and the organization slows down precisely as the market demands it move faster. The old military line, let the people who hear the gunfire make the decisions, captures the antidote: push decision rights to the edges where the information is freshest.
Organizational architecture frames this as a matching problem. Decisions should be made where the relevant information is richest and most current. At the front line, knowledge of the customer and the immediate problem is abundant and fresh. At the top, that same knowledge arrives filtered, delayed, and degraded through layers of summarization, which is why scaled organizations make slower and often worse operational calls than the scrappy startups they used to be.
The counterweight the decentralization evangelists underplay is the principal-agent problem. Pushing decision rights to the edges assumes the edges will decide in the organization's interest, and that assumption collapses when incentives are misaligned or local decisions carry non-local consequences. A front-line team optimizing for its own customer can make a choice that is locally brilliant and globally destructive, a discount that wins the deal and quietly breaks the pricing model, a feature that thrills one client and fragments the product for everyone else. Pure decentralization without aligned incentives and clear boundaries doesn't produce agility. It produces a thousand locally rational decisions that sum to strategic incoherence and a very confusing all-hands.
The resolution isn't choosing between centralized and decentralized. It's being deliberate about which decisions belong where. Reversible, local, information-rich-at-the-edge decisions should be decentralized aggressively, because the cost of a wrong call is low and the value of speed is high. Irreversible, global, or centre-only-visible decisions should stay centralized. The organizations that keep their agility at scale draw this line explicitly and push everything they safely can toward the gunfire, while holding the genuinely strategic and irreversible choices at the centre. The failure mode is never simply too much or too little decentralization. It's failing to tell which decision is which, and applying one blunt rule to both.