Adaptive Orchestration with Cross-Episode Memory for Dynamic LLM-based Agent Pools

Abstract

Multi-agent systems are moving toward open ecosystems in which agents may appear, change, or disappear at runtime. Orchestra-tors must therefore discover relevant black-box specialists from large pools and learn how to coordinate them across repeated tasks. Existing methods typically either maintain cross-episode memory for fixed, well-known teams or retrieve agents without accumulating reusable coordination knowledge. We propose an adaptive orchestration architecture with an exter-nal memory layer that enables in-context learning over dynamic agent pools without parameter updates. This layer combines agent play-books, which store per-agent evidence, with delegation blueprints, which capture reusable workflows distilled from successful episodes. On OfficeBench with a 50-agent pool including distractors, we show that the adaptive sys-tem achieves the strongest overall performance across all evaluated configurations when operating with its own Agent Cards, outperforming both sparse-card and rich-card stateless baselines while requiring fewer orchestrator turns. Ablations show that blueprints drive the main end-to-end gains, whereas playbooks are most useful when descriptions are sparse. A dynamic-pool experiment further shows that the memory layer incorporates newly added specialists and prunes obsolete capability evidence without manual intervention.