Your agent's world model is only as fresh as your RPC latency. SVS is bare-metal infrastructure tuned for autonomous workloads: flat-rate pricing, sub-millisecond account queries, and co-located compute on the same network as the validator.
Most Solana RPC stacks were built for backends with predictable QPS. Agents are not predictable. They poll, react, retry, and run hot for hours. The economics and the latency profile fight them at every step.
Five primitives, all on the same physical network. Pick one or run the whole stack. Colocated compute is the multiplier that makes the rest work.
Or it doesn’t, and the spread you saw belongs to someone else. The infrastructure decision compounds across every cycle of every loop.