Governed Agency: A Non-Confidential Synthesis of Controlled Experiments in Agent Construction, Recursive Adaptation, Multi-Agent Coordination, Observability, and Governance

Governed Agency is a non-peer-reviewed, claim-bounded synthesis of controlled experiments, engineering prototypes, negative results, and governance architectures across 133 project roots. The report examines how effective agent capability is assembled from a language model, memory, tools, persistent state, shared environments, evaluators, and authority, and why governance must address the capability lifecycle rather than isolated model outputs. The documented evidence includes bounded results on directed relay, anonymous shared-state coordination, persistent external artifacts, recursive system-state learning, governed policy search, path-dependent observability, capability controls, rollback, and trust boundaries. It also preserves failed or unsupported claims concerning stable emergent roles, shared language, open-ended recursive self-improvement, autonomous goals, provider-side weight learning, universal detection, production-grade containment, and independent trust on a single host. The report is non-confidential and deliberately excludes private prompts, source code, credentials, detailed attack recipes, customer configurations, proprietary scoring logic, and reconstructable implementation thresholds. Evidence statements are restricted to their recorded protocols. This version is a preprint and has not been peer reviewed.