The same foggy maze, solved by two agent-team topologies built on sagent primitives. A lock opens only when two different agents press its two far-apart plates, each naming the other, within a short window — so the team must spawn, explore, and coordinate by message. Decentralized: any agent spawns helpers and messages anyone. Centralized: only the coordinator spawns, and every message relays through it.
Counts look inverted on purpose: decentralized logs more total messages because it completes all three coordinations; centralized logs fewer only because it never finishes one. Read cost per lock, not the raw totals — a stuck arm has no per-lock cost at all.
One row per agent (a0 = the first agent, on top). Each is born at its dot; a vertical drop links a parent to each child it spawns. Decentralized: spawned agents spawn more, so the tree branches. Centralized: only the coordinator ever spawns — one lane, a flat star.
Each lock is an independent two-agent rendezvous: find your partner's identity, agree the pairing, and co-press. Over any-to-any messaging each of those is one hop, and locks are coordinated in parallel. Through a single hub every fact double-hops and the coordinator serializes K independent coordinations — so the cost-per-lock climbs and, when the topology is hidden, the centralized team often never converges at all. The headline below is coordination cost per lock opened (delivered messages ÷ locks opened); a stuck arm reads as high cost at a low solved fraction, never as a tie. This is also why the decentralized arm can show more total messages than the stuck centralized one — it actually finishes all three coordinations, where the centralized one finishes none.