Rule-based elision before summarization led on efficiency among context-management strategies

Trajectory analysis is where the three components separate. Context management extends execution trajectories without substantially altering agent behavior ^[14], planning changes where trajectories stop ^[15], and the action space changes the granularity at which code is written ^[16]. Most of context management's value comes from preventing context-overflow failures, which otherwise end a run before the agent gets to code modification and verification ^[4][6].

The elision result is narrower than the ranking suggests. Staging rule-based elision ahead of LLM-based summarization was the strongest of the five strategies on overall efficiency ^[7], and the paper describes it as maintaining "mean success similar to the other managed strategies while controlling peak context and reducing reliance on summarization calls" ^[8]. Equal accuracy for fewer summarization calls is a cost finding. The more elaborate variant, in which elided content can be recovered, "adds machinery that models rarely use and yields no accuracy gain", the authors write ^[9].

The planning result splits by model in a way that matters for anyone reading a single benchmark number. For weaker models, planning keeps the trajectory alive long enough to attempt an edit, raising success at additional cost; for stronger models it mainly removes redundant post-edit verification, lowering cost ^[11]. The authors frame the contribution as informing model- and budget-aware harness design ^[19], which is a fair description of a set of findings that change sign depending on the model in the loop.

Five context-management strategies at four context-window budgets on four models comes to 80 settings, leaving 96 for the targeted planning and action-space ablations ^[21]. The paper page gives the abstract and a findings summary; it does not name the four models or report absolute success rates and costs ^[20]. The posted summary states the gap the study is aimed at: "Most evaluations compare complete agent systems, making it difficult to separate the contributions of individual harness components" ^[18].

Four unnamed models on two benchmarks is a thin base for point estimates. The direction holds across all four: weak-model and strong-model configurations diverge, and they diverge mostly in cost ^[11][12].

For a team already running an agent, the ordering in the paper is the usable part: overflow prevention is what context management mostly buys back ^[4], and the recoverable variant bought no accuracy at all ^[9].