GitHub found that compressing tool output made Copilot tasks more expensive, not less

GitHub Blog·high signalXBlueskyLinkedInCopy link

GitHub published four measured cost reductions for Copilot on September 2: selective output compaction 5.5%, removing view-tool line-number prefixes 3.1%, a compacted task-tool prompt 2.9%, and batching background-task results 2.3%. The counterintuitive result came from testing RTK (Rust Token Killer), which shortens shell output: individual responses got smaller but tasks got more expensive, because the agent reopened the original output or reran the command to recover what was cut. GitHub's conclusion is a directly reusable rule for anyone tuning an agent loop: "tokens per tool call is the wrong objective" because cost has to be measured over the whole task.