Agentic Evolution: From Self-Improving Agents to Co-Evolving Human-AI Systems

Agentic evolution concerns not just how agents improve, but whether the signal driving that improvement remains reliable as the system evolves. This survey organizes roughly 300 papers on agentic evolution through a three-axis taxonomy: evolutionary substrate, consolidation pathway, and selective pressure. When read across the first two axes, these papers yield regularities invisible within any single axis: consolidation failure modes track the pathway rather than the substrate, and the pathway choice is jointly constrained by artifact discreteness, evaluation-signal verifiability, and infrastructure access. The third axis surfaces a distinction that prior surveys, to our knowledge, omit or subsume: whether evolution is shaped by autonomous or human-involved selective pressure. Autonomous evolution produces its strongest reported results where deterministic verifiers, independent of the system being evaluated, are available; absent such verifiers, self-referential and proxy-based signals yield diminishing returns and can degrade with iteration. Beyond this boundary, the evidence suggests reliable evolution depends on human-involved selective pressure. Yet such pressure is rare, low-bandwidth, and not the fixed input that current systems assume.

A reverse analysis of nearly 100 studies from cognitive science, education, and labor economics finds converging evidence that default, unscaffolded AI interaction can reduce the independent evaluative capacity on which such pressure depends. This effect is not universal; it is bounded by task structure and interaction design, and often undetected by users. Together, these literatures motivate reinterpreting the third axis: human input, originally modeled as exogenous, is better understood as endogenous, shaped by the process it is meant to guide. We propose treating agentic evolution and human adaptation as a co-evolving system, and outline a research agenda for monitoring and maintaining both partners’ capacities across deployment lifetimes.