Before AI Agents Evolve in the Wild - Pre-Deployment Evolutionary Stress Tests of AI-Agent Populations with the PVPP Framework Motivated by the 2026 OpenAI–Hugging Face Incident

What happens when AI agents do more than act once—when they persist, share information, inherit configurations, use tools, accumulate resources, and change the environment faced by later agents? This white paper develops a pre-deployment stress-testing approach for those population-level dynamics using the Productive Value–Productive Power (PVPP) framework. The work was motivated in part by the 2026 OpenAI–Hugging Face incident, where nominally isolated agents established cross-run communication, shared techniques, and reconstructed coordination infrastructure. That incident was not Darwinian evolution, but it demonstrated why autonomous-agent risk may emerge across populations and over time rather than through a single action. The paper reports a staged experimental program ranging from reproducible controlled ecologies to external data and live LLM execution. Among the main findings: controlled agent populations demonstrated generational selection, environment-specific inherited adaptation, and reciprocal agent-agent coevolution without defining Productive Power as evolutionary fitness or using a global fitness ranker; a purpose-built model-backed system demonstrated a complete downstream pathway from environment-specific operability to local resources, reproduction, and lineage success; a live permission-gated agent population reproduced much of that pathway with fresh LLM execution, passing six of seven preregistered gates; analysis of the TerraLingua agent ecology showed that strongly inherited configuration did not automatically produce inherited measured capability, an important distinction for cloned, forked, or modified agents; two earlier live-agent experiments were stopped at their qualification gates because nominal tool possession did not translate into reliable tool use, demonstrating the value of testing the execution interface before drawing population-level conclusions; a stronger hypothesis of reciprocal agent-information coevolution was not established under the final corrected control and was retained as a negative result. The paper does not claim that deployed AI agents generally evolve, nor that the PVPP framework predicts arbitrary real-world deployments. Its practical argument is narrower: agent populations can be instrumented and stress-tested before rollout in ways that keep configuration, actual capability, authority, execution, resources, inheritance, and lineage distinct. The accompanying reproducibility archive includes the controlled-stage source code, outputs, checksums, analysis materials, and documented evidentiary gaps.