Two Evolutionary Trajectories and a Process-Regulation Gap: From the Evolution of Agency and External Trace to Large Language Models and AWM

Large language models combine a high-capacity computational substrate with access to a large portion of digitized human linguistic and cultural trace. This working paper develops a conceptual account of that convergence through two long trajectories: the evolution of agency and action capacity, and the evolution of externalized trace, transmission, and cumulative information. Their convergence in LLM-mediated interaction expands human access to knowledge and generative transformation, but also creates a process-regulation gap concerning goals, provenance, epistemic status, execution state, responsibility, and reality contact. The argument is grounded in three previously published naturalistic case studies: Tool-Assisted Sycophancy, Recursive Sycophancy, and False Execution Status. AWM (AI Working Method) is presented as a case-grounded process architecture that emerged through longitudinal human–LLM work. Its proposed function is homeostatic rather than generative: to coordinate the human operator, the LLM, external memory, tools, and reality feedback while controlling process transitions. The paper distinguishes observed case evidence from broader hypotheses and proposes a research programme for external validation.