Shrinking an AI agent's context window can look free on the scoreboard while quietly loading up its workload elsewhere.
A new arXiv paper builds a controlled test rig to measure what happens when a tool-using AI agent has its context trimmed mid-task. Researchers ran three models through a deterministic 24-turn planning environment, comparing a compression method that simply drops old state against one that tries to preserve key facts, then split every tool call into retrieval versus execution. Across all six model-and-task combinations tested, retrieval calls rose after compression, and that rise accounted for nearly all the added interaction. Task completion itself barely budged: GPT-5.5's success rate moved from 80 percent to 85 percent after heavy compression, a statistically meaningless shift, even as its retrieval calls jumped from 21 to nearly 64.
Most benchmarks for agentic AI still score a run by whether the task finished, the same blind spot this paper targets. That framing hides a real cost: an agent making three times as many retrieval calls to reconstruct facts it already had is burning latency and API spend that never shows up in a completion percentage. The effect was not universal either, since a second test in the ALFWorld environment showed no retrieval surge at all, meaning the hidden tax depends on how a given environment handles state, not on compression itself.
If your agent framework brags about a smaller context footprint, ask what it did to your tool-call bill before you believe the savings.