AI agents may be able to complete many extended workflows without carrying their entire interaction history into every model call. A November 3, 2026 arXiv preprint by Peter Belcak and colleagues reports that repeatedly compacting that history largely preserved benchmark accuracy while keeping working contexts shorter. Peer review is not verified, and the supplied evidence includes no independent replication.
The study compared full-history accumulation with a process that asks the model to produce a compact account of task-relevant state once a token threshold is reached. A fresh conversation then receives the original task and that account. These compact records emphasize requirements, current work, next steps and relevant artifacts rather than simply recounting past actions.
The main experiments used Claude Opus 4.5 and GPT 5.6 Sol across six long-horizon benchmarks covering software repairs, terminal tasks, information retrieval and tool use. For coding, terminal and FRAMES tasks, tested total context limits ranged from 48,000 down to 16,000 tokens; some tool-use settings used smaller limits. Evaluation relied on repository tests, task verifiers or benchmark-specific checks, with an LM judge used for FRAMES.
Across 22 compaction configurations per model, accuracy changes relative to full-history baselines ranged from −3.7 to +5.6 percentage points for Claude and −3.3 to +4.4 points for GPT. For each model, no more than five configurations lost over two percentage points. These results suggest that frequent compaction can preserve much of the tested performance, but small differences alone do not establish statistical equivalence.
Separate experiments measured open-weight models on fixed H100 GPU deployments. With cached input prefixes and saturated serving, reducing configured context size from 256,000 to 16,000 tokens increased peak decoding throughput by 3.6-fold for GLM 5.2, 5.6-fold for Kimi K2.6 and sevenfold for Qwen 3 Coder Next. These measurements concern output-token generation under selected serving conditions, rather than complete workflows including input processing, compaction and tool execution.
The authors also estimated long-horizon cost reductions of 2.2–8.8-fold for Claude and 2.4–5.0-fold for GPT. Those are modeled savings, not measured reductions in customers’ API bills; the supplied cost-model appendix is incomplete.
For long-input tasks, the researchers tested selective reading of documents stored in a filesystem across seven benchmarks, with compaction available when needed. Because this approach changes document access and tool use alongside context management, its effects cannot be attributed to compaction alone. External files also remain an information store: a short active context does not mean the system retains little information overall.
Compaction can discard details needed later. Task-level independence, benchmark contamination and selection effects across many comparisons are not established, and the supplied results do not provide confidence intervals or statistical equivalence tests. The evidence supports shorter working contexts for many tested agent workflows, while leaving open when long-context capability remains beneficial.
Source: https://arxiv.org/html/2611.05306v1
Cover illustration generated with AI; conceptual, not research evidence.