Research / CGP-V1
GovernancePublished · June 2026

CGP: Cyryx Governance Protocol for Agentic AI Execution

Version 1.0 — Technical Report CGP-2026-001

Authors: Cyryx Labs LLCDate: June 2026DOI: 10.5281/zenodo.21045760Concept DOI (all versions): 10.5281/zenodo.21045759License: CC BY 4.0

Abstract

Existing AI governance frameworks — the EU AI Act, the NIST AI Risk Management Framework (AI RMF), and ISO/IEC 42001 — were designed for AI systems operating under continuous human supervision: classifiers, recommenders, and single-turn generators. None were designed for agentic AI systems that autonomously decompose goals into multi-step plans, execute sequences of environment-modifying actions, coordinate multiple specialized sub-agents, and maintain state across sessions. Singapore's Model AI Governance Framework (January 2026) is the only published governance document that acknowledges this gap, identifying three unaddressed risks: cascading failure propagation, emergent scope expansion, and attribution gaps across agent chains.

This document introduces the Cyryx Governance Protocol (CGP) v1.0, a technical framework that fills these gaps with seven control domains and twenty-eight normative controls (MUST/SHOULD/MAY). CGP is designed as an extension to existing frameworks — not a replacement — with explicit mapping to EU AI Act Articles 9, 12, 13, 14, and 15; NIST AI RMF functions GOVERN, MAP, MEASURE, and MANAGE; and ISO 42001 Clause 6, 7, 8, and 9 controls. Every control in CGP v1.0 has a reference implementation in MAAX Studio by Cyryx Labs. CGP is published under Creative Commons Attribution 4.0 (CC BY 4.0) for open community adoption and review.

Keywords:agentic AIAI governanceEU AI ActNIST AI RMFISO 42001autonomous agentsmulti-agent systemsAI safetycompliance frameworkgoverned execution

The Three Gaps CGP Fills

Cascading Failure Prevention

When an agent fails mid-execution, failures propagate silently. CGP enforces rollback checkpoints and gate-failure halts.

Scope Creep Control

Agents expand their own scope without authorization. CGP enforces explicit scope boundaries that cannot be exceeded implicitly.

Decision Attribution

"The AI decided" is not auditable. CGP requires operator-level, model-level, and human-level attribution for every consequential decision.

Seven Control Domains · 28 Normative Controls

CD1

Mission Authorization

4 controls · 3 MUST

CD2

Scope Boundary Enforcement

4 controls · 4 MUST

CD3

Checkpoint Integrity

4 controls · 3 MUST

CD4

Command Gate System

4 controls · 3 MUST

CD5

Decision Attribution

4 controls · 4 MUST

CD6

Evidence Chain

4 controls · 4 MUST

CD7

Cost Governance

5 controls · 4 MUST

Mapping to Existing Frameworks

CGP extends — not replaces — the EU AI Act, NIST AI RMF, and ISO 42001.

CGP DomainEU AI ActNIST AI RMFISO 42001
CD1 — Mission AuthorizationArt. 14, Art. 9GOVERN 1.1, MAP 1.5Clause 6.1, Annex A-6.2
CD2 — Scope BoundaryArt. 9, Art. 13MAP 3.5, MEASURE 2.5Annex A-8.4, Clause 8.1
CD3 — Checkpoint IntegrityArt. 9.4MANAGE 2.2, MANAGE 3.1Annex A-8.5, Annex A-9.1
CD4 — Command GatesArt. 9, Art. 15MEASURE 2.1, MEASURE 2.3Annex A-8.3, Annex A-8.6
CD5 — Decision AttributionArt. 12, Art. 14.4GOVERN 6.1, GOVERN 6.2Annex A-8.2, Clause 7.5
CD6 — Evidence ChainArt. 12, Art. 17GOVERN 6.2, MEASURE 2.8Clause 9.1, Annex A-8.7
CD7 — Cost GovernanceNot explicitly addressedGOVERN 4.1, MANAGE 4.1Annex A-8.4, Clause 9.1

Conformance Levels

BASELINE

CGP-L1

Agentic Baseline

All MUST controls in CD1, CD3, CD5, CD6

GOVERNED

CGP-L2

Governed Execution

All MUST controls across all 7 domains

ENTERPRISE

CGP-L3

Enterprise Agentic Governance

All MUST and SHOULD controls + third-party review

CGP v1.0 is not yet a certifiable standard. Organizations may claim "CGP v1.0 conformant at Level [1/2/3]" after self-assessment.

Reference Implementation

Every control in CGP v1.0 has a reference implementation in MAAX Studio by Cyryx Labs. MAAX Studio is the first IDE built around governed agentic execution — with Mission State Machine, Command Gates, Evidence Vault, and Secure Apply Engine as architectural primitives.

Learn about MAAX Studio →

Implemented in MAAX Studio

  • Mission State Machine (14 states, 20 transitions)
  • Command Gates (6 deterministic quality gates)
  • Evidence Vault (cryptographic audit chain)
  • Secure Apply Engine (transactional rollback)
  • Mission Ledger (real-time cost attribution)
  • Operator Corps (7 specialized operators)

Contribute to CGP

CGP is published under CC BY 4.0. Open an issue, submit a gap report, or share your implementation.

View on GitHub →

Contact

Cyryx Labs LLC, United States of America