Constitutional Self-Governance for Autonomous AI Agents: A Framework Observed in 77 Days of Production
Year: 2026
Type: preprint
Abstract: As organizations deploy autonomous AI agent systems, a critical governance gap has emerged. Current approaches govern agent permissions — identity, access control, audit logging — but not agent behavior. We term this the "WHO vs. HOW" gap. This paper presents Constitutional Self-Governance (CSG), a framework that constrains autonomous agent decisions through 12 interlocking mechanisms organized into structural, operational, and evolutionary tiers. The framework has been observed in continuous production for 85 days with 48 registered agents, 60 constitutional amendments, three P0 incidents self-resolved, and less than 30 minutes of daily human oversight. The system reports its own failures honestly — currently operating in FREEZE state ($0 revenue) rather than fabricating success metrics. Aligns with EU AI Act (August 2026), NIST AI Agent Standards, Singapore Model AI Governance Framework, and OWASP ASI Top 10 (8/10 compliance). Companion to the Decision Load Index preprint (DOI: 10.5281/zenodo.18217577). v1.3.0 (2026-09-05): production metrics updated from Day 77 to Day 85 (48 registered agents after registry consolidation); adds structured harness validation of OWASP ASI compliance. Supersedes v1.2.1 (2026-03-22), which reported the 77-day window with 56 agents and remains the correct citation for that window.
Author: Saleme, Michael K.
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7140035835