Coding agents routinely receive tool output that is too large for a single turn, yet a study of public Model Context Protocol session logs found no agent-initiated request for a second chunk. Researchers tested whether putting the most useful file first would improve performance.
They treated first-chunk selection as an optimization problem and compared six ranking methods on 500 SWE-bench Verified tasks. A keyword scorer raised the rate of placing the needed item first from 24.2% to 35.0%, or 35.8% when it fell back to the tool’s original ordering.
That ranking gain did not reliably improve a separate file-localization test across five language models and 4,800 calls. Changes stayed below three percentage points, varied in direction, and were not statistically significant. The agents could recover the required item from anywhere in the visible chunk, making inclusion more important than first position.
This was a single-turn localization probe, not an end-to-end test of whether agents fixed software bugs. The result cautions tool designers against optimizing an easy ranking metric without checking downstream behavior. Better chunk selection may mean maximizing relevant coverage, while agents themselves still need more dependable strategies for requesting omitted results.