Externalization in LLM Agents (arxiv.org)

A recent paper titled "Externalization in LLM Agents" explores the evolving methodology of developing large language model (LLM) agents, emphasizing a shift from internal model adjustments to the integration of external cognitive tools. The paper reviews how capabilities traditionally embedded within models are now externalized into memory systems, reusable skills, and structured protocols, collectively enhancing the reliability and functionality of LLM agents. This approach is grounded in the concept of cognitive artifacts, illustrating that a well-designed agent infrastructure can alleviate cognitive demands, making processes more efficient. This review is significant for the AI/ML community as it proposes a systems-level framework that highlights the importance of external cognitive infrastructure in advancing agent capabilities. By categorizing memory, skills, and protocols as distinct yet interconnected forms of externalization, it opens up avenues for future research, including self-evolving harnesses and the shared infrastructure between agents. Crucially, it identifies the trade-offs between parametric costs and externalized capabilities, urging the community to focus not just on developing more powerful models, but also on creating cohesive and effective external systems that support them.