Agent requests grow between turns
A coding-agent request can contain system instructions, conversation history, repository context, tool definitions, and prior tool results. Much of that envelope may be sent again on every turn.
The task itself can be short while the surrounding payload dominates the billable input.
Efficiency is selective, not destructive
Payload optimization should not blindly truncate the request. The active goal, relevant evidence, required protocol, and forced tool choice must survive filtering.
Eligible noise includes stale conversational turns, duplicate instructions, oversized completed tool output, and schemas for tools that are not useful to the current step.
Measure retained context
A useful system records what entered the router and what remained after policy was applied. That makes token reduction visible without exposing every internal heuristic.
Payload efficiency works alongside routing and provider caching. Keeping those effects separate makes cost evidence easier to interpret and prevents one optimization from taking credit for another.