SJTU and 21 Others Publish Integrated Review Paper on Externalization of LLM Agents

• Argues that the development of LLM agents can be explained by 'externalization,' which shifts cognitive load to memory, skills, protocols, and harnesses, rather than by modifying model weights.
• Proposes the concept of 'representational transformation,' based on Donald Norman's theory of cognitive artifacts, suggesting that external tools change the very shape of the task.
• Classifies memory as temporal state (recall → re-presentation), skills as procedural expertise (ad-hoc generation → verified combinations), and protocols as interaction structure (improvisation → structured contracts).
• Defines the harness not as a fourth type of externalization, but as the execution environment and engineering layer in which the previous three operate.

Roles by Externalization Type

• Memory: Temporal state management that resolves issues with finite context windows and the lack of inter-session memory.
• Skills: Verified component combinations that resolve consistency issues (variance) in long multi-step procedures.
• Protocols: Structured contracts that resolve instability in interactions with external tools and agents.

Skill Externalization Stages and Safety

• Composed of five stages: specification, exploration, progressive disclosure, execution binding, and composition; during progressive disclosure, Level 1 uses approximately 100 tokens, and Level 2 uses less than 5k tokens.
• Vulnerabilities such as prompt injection were found in 26.1% of 31,132 skills, and skills containing execution scripts have a 2.12 times higher probability of vulnerabilities.
• When applying curated skills based on SkillsBench, the average pass rate improved from 33.9% to 50.5%.