Green SARC: Predictive Cost and Carbon Governance for Agentic AI Systems

The article introduces Green SARC, a governance framework for managing the financial and environmental costs of agentic AI systems through architectural enforcement in the agent loop. It reports key findings, including a $\Theta(n^2)$ complexity for the "State Snowball" model, a median curvature of 216 in real multi-step plans, and an effective soft Lagrangian penalty that breaches budget expectations in 91.5% of cases, while maintaining a 0% over-budget incidence with architectural gates. This framework is significant for practitioners as it provides a structured approach to predict and control costs in AI deployments, emphasizing the importance of policy-independent results and open-source accessibility for further development.