Autonomous agents tackling multi-hour workflows routinely crash into a familiar wall: context exhaustion. Standard workarounds—whether blunt text compression or lossy RAG lookups—inevitably discard low-level tool outputs long before downstream reasoning steps actually need them. As Yin Lin of Alibaba Group, Elaine Ang of Columbia University, and their co-authors outline in an arXiv preprint, the root problem is treating context as a static prompt dump rather than an active programmatic runtime.
Their proposed architecture, Scroll, replaces static conversation history with a stateful execution environment. At its foundation sits an append-only Event Log that preserves exact operational history, paired with a persistent, sandboxed Python kernel. Rather than serializing massive JSON payloads or tool dumps straight into working context, Scroll assigns outputs to typed variables. The model manipulates state via code execution, projecting only explicitly rendered results into its active prompt view. When working capacity shrinks, stale blocks are offloaded to an index structured with compact landmark pointers, giving the agent deterministic random access to raw operational logs without prompt bloat.
Evaluating the architecture with Qwen3.8-Max as the backbone demonstrates the practical advantage over naive memory pooling. Scroll hit 94.8% on LongMemEval S and 73.1% on BEAM10M, beating the top published memory baseline by 5.1 points. On the LOCA256K benchmark for extended horizons, it logged 86.7%—a 37.4-point leap over previous systems. For enterprise teams running autonomous multi-step pipelines, shifting from prompt serialisation to programmatic state management looks like the most viable path to maintaining reasoning accuracy without torching token budgets.