When Does Context Compaction Pay in Long-Horizon Agents?

Abstract Compacting an agent's history shortens later requests but can change the work needed to finish its task. We ask when early compaction pays after counting that change: only when a fixed-trace savings budget exceeds the cost of the changed work. We measure both with matched-state continuations to termination on nine long Terminal-Bench tasks, complete runs in two agent runtimes, and replays on a dedicated GPU deployment. At a 64K threshold, gpt-6.1-sol's budget was USD 0.17 per state, yet synchronous compaction (Sync) cost USD 0.26 more with measured cache reads. Added requests, mostly reads, exhausted the budget; the excess persisted across prompt, task, and runtime controls. With GLM-5.3-Flash, compaction saved. Cache prices set the budget: for our nine-task mix of complete 64K Sol runs, Sync would cost less only above a cache-read price of 34% of the uncached price, against a listed 5%. At one common price vector, savings appeared only where much work remained. On a dedicated GLM deployment, replayed Sync sessions took 30% less mean time than sessions keeping the full history at light load (3% with the work fixed) and at least 29% less at heavy load. At 102 concurrent sessions without a host-memory cache tier, waiting sessions' evicted prefixes had to be computed again, and Sync's work with the history restored took more than twice as long. We release an offline calibrator mapping complete-task usage and deployment prices to policy costs, price crossovers, and selection headroom.