Synthetic Chaos as an Institutional Laboratory: Independent Evidence for Extended Phenotype Theory from Autonomous Agent Red-teaming

The study Agents of Chaos (Shapira et al., 2026, arXiv:2602.20021) documents eleven failure cases in autonomous agent systems with persistent memory, tool access, and multi-party communication. The AI safety community reads those failures as urgent security problems; this article offers a distinct and complementary interpretation: they constitute independent empirical confirmation of the central predictions of Extended Phenotype Theory (EPT) applied to law. Five convergences are identified between AoC findings and the EPT/EGT research program: local alignment without global stability, memetic replication through instruction invasion, horizontal entrenchment via echo chambers, absence of group-level optimization, and an administrative ratchet effect. On that foundation, the article develops its more ambitious thesis: synthetic agent populations with evolutionary time represent the first genuine experimental laboratory for legal institutional design. Unlike classical agent-based modeling, LLM-powered agents combine learning capacity, persistent memory, and cultural transmission, three conditions that enable hypothesis testing about institutional arrangements before deploying them on real populations. Five operational tools are described (JurisRank, RootFinder, HisteresisCalc, IusMorfos, LLM Council), together with a five-step experimental protocol and validation data across 60 historical reforms. One verified case study, with six months of post-implementation follow-up data, demonstrates that evolutionary simulation identified a cross-domain evasion strategy that traditional legal analysis did not anticipate. Two illustrative scenarios, derived from the same mechanisms and presented as falsifiable model predictions, extend the methodology to corporate compliance and tax reform contexts. The synthetic laboratory is not five years away: it is operational, documented, and accessible.