GitHub measures that shortening agent tool output raised total task cost in its own harness

Sources: primary, discussion
Summary: GitHub evaluated RTK, a utility that truncates shell output before the agent reads it, and reports that when the omitted text mattered the model sometimes reopened the original output or reran the command to recover what it needed, so the task on average used more tokens and took longer. GitHub scopes the result to the integration and workloads it tested rather than to output compression in general. It states the four changes it did ship were picked offline on agentic coding benchmarks and then validated in controlled online A/B experiments.
Why it matters: Shortening each tool response is a common way to cut agent spend, and GitHub publishes a measurement in which that move raised the cost of the whole task rather than lowering it.
Follow-up: Track whether GitHub publishes the harness or the workload set behind the RTK result.