Context Kubernetes: Declarative Orchestration of Enterprise Knowledge for Agentic AI Systems
Charafeddine Mouzouni · 2026
Read Context Kubernetes: Declarative Orchestration of Enterprise Knowledge for Agentic AI Systems by Charafeddine Mouzouni (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-11623.txt.
Introduction
Mouzouni introduces Context Kubernetes, a design for running enterprise knowledge for agentic AI. The core analogy: delivering the right knowledge to the right agent with the right permissions and freshness resembles container orchestration. Six abstractions, a YAML manifest, a sync loop, and a three-tier permission model keep agent authority below human authority. Against four baselines, ACL filtering stops cross-domain leaks, intent routing cuts noise by 19 points, and only the three-tier model blocks all five tested attacks. TLA+ checking verifies safety across 4.6 million states with zero violations.
Catalog record
| CREATOR | Charafeddine Mouzouni |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66495 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.11623
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Context Kubernetes: Declarative Orchestration of Enterprise Knowledge for Agentic AI Systems Charafeddine Mouzouni Abstract . Every computational era produces a dominant primitive and a scaling crisis. Virtual machines needed VMware. Containers needed Kubernetes. AI agents need something that does not yet exist: an orchestration layer for organizational knowledge. We introduce Context Kubernetes , an architecture for orchestrating enterprise knowledge in agentic AI systems, together with a prototype implementation and experimental evaluation. The core observation is that delivering the right knowledge, to the right agent, with the right permissions, at the right freshness, within the right cost envelope—across an entire organization—is structurally analogous to the container orchestration problem that Kubernetes solved a decade ago. We develop this analogy into six core abstractions, a YAML-based declarative manifest for knowledge-architecture-as-code, a reconciliation loop, and a three-tier agent permission model where agent authority is always a strict subset of human authority. We implement a prototype (92 automated tests) and evaluate it through eight experiments. Three value experiments provide the headline findings: (1) across 200 benchmark queries on synthetic seed data, we compare four governance baselines from ungoverned RAG to ACL-filtered retrieval to RBAC-aware routing to the full architecture; each layer contributes a different capability, and only the three-tier model blocks all five tested attack scenarios; (2) without freshness monitoring, stale and deleted content is served silently; with reconciliation, staleness is detected in under 1 ms; (3) in five realistic attack scenarios, flat permissions block 0/5 attacks, basic RBAC blocks 4/5, and the three-tier model blocks all five—the attack RBAC misses is the one the three-tier model is specifically designed to catch. Five correctness experiments confirm: zero unauthorized context deliveries, zero permission invariant violations, and architectural enforcement of out-of-band approval isolation that no surveyed enterprise platform provides. TLA+ model-checking verifies the safety properties across 4.6 million reachable states with zero violations. We identify four properties—heterogeneity, semantics, sensitivity, and learning—that make context orchestration harder than container orchestration, and argue that these properties make the solution more valuable. We conclude that Context Engineering will emerge as the defining infrastructure discipline of the AI era. 1. Introduction The most profound technologies are those that disappear. They weave themselves into the fabric of everyday life until they are indistinguishable from it. Mark Weiser , 1991 Every generation of computing infrastructure follows the same arc. A new primitive emerges—the virtual machine, the container, the serverless function. Pioneers demonstrate extraordinary results on a single machine. Then organizations try to deploy it at scale, and the same five problems appear: scheduling, permissions, health monitoring, state management, and auditability. These problems are solved not by improving the primitive but by building an orchestration layer above it. VMware orchestrated virtual machines. Kubernetes orchestrated containers. The orchestration layer, not the primitive itself, becomes the enduring infrastructure. The AI agent is the new primitive. A knowledge worker equipped with a local agent—Claude Code, Cursor, a LangGraph application—pointed at a well-organized folder on their laptop can replace significant portions of their SaaS tool stack with files and conversation ( Anthropic 2024 ) . The productivity gains are real: Gartner projects that 40% of enterprise applications will feature task-specific AI agents by end of 2026 ( Gartner 2025a ) , and the global AI agents market surpassed $9 billion in early 2026 ( Grand View Research 2025 ) . The primitive works. The scaling crisis has already begun. Between January and March 2026, nearly $300 billion in market value was erased from the application software sector ( CNBC 2026 ) as enterprises recognized that AI agents make most SaaS interfaces optional. But going from one agent on one laptop to 2,000 agents across an organization immediately surfaces the same five problems: • Scheduling : Which knowledge reaches which agent, from which source, in what order? • Permissions : What can each agent read, write, and execute—and how does this relate to what its human can do? • Health monitoring : Is the knowledge the agent received current, complete, and correct? • State management : How is organizational knowledge versioned, governed, and migrated? • Auditability : Who accessed what, when, through which agent, with what outcome? Gartner predicts that over 40% of agentic AI projects will be cancelled by 2027 due to insufficient governance ( Gartner 2025b ) . The technology is mature. The orchestration layer is missing. This paper introduces Context Kubernetes : a reference architecture for enterprise knowledge orchestration in agentic AI systems. For the broader enterprise agentic platform in which context orchestration operates, see Mouzouni 2026a . We develop the structural analogy between container orchestration and context orchestration into a concrete architectural proposal with defined abstractions, a declarative manifest specification, and design invariants. The analogy is productive—every major Kubernetes primitive has a functional counterpart in the knowledge domain—but it is an analogy, not a formal isomorphism, and we are explicit about where it holds tightly and where it stretches. Scope. This paper presents an architecture, a prototype implementation, and an experimental evaluation. The architecture defines the abstractions and design invariants. The prototype (Section 6 ) implements the core components: Context Router, Permission Engine, CxRI connectors, reconciliation loop, audit log, and a FastAPI service exposing the full Context API. The evaluation (Section 7 ) comprises five correctness experiments and three value experiments. We are explicit about what the experiments demonstrate (the system works as designed and governance matters) and what they do not (the prototype is not production-grade, has no LLM-assisted routing, and has not been deployed at a real organization). 1.1. The Orchestration Emergence Thesis We propose the following general thesis: When a new computational primitive reaches organizational scale, the orchestration layer that governs its lifecycle—scheduling, permissions, health, state, and audit—becomes the most valuable and enduring infrastructure layer, exceeding the value of the primitive itself. Historical evidence supports this thesis. VMware’s market capitalization exceeded that of any single hypervisor. Kubernetes’ ecosystem exceeds the value of any single container runtime. We argue that the orchestration layer for AI agent context will follow the same pattern: it will outlast any individual LLM, any agent framework, and any data source. LLMs will be replaced. Frameworks will evolve. The orchestration layer persists. 1.2. Contributions This paper makes six contributions: (1) A reference architecture with six core abstractions, a declarative manifest specification, and a structural analogy to Kubernetes identifying where the mapping is tight, moderate, and loose (Sections 3.2 – 5 ). (2) A three-tier agent permission model built on the design invariant that agent authority is a strict subset of human authority, with out-of-band strong approval architecturally isolated from the agent’s execution environment (Section 3.4 ). (3) A prototype implementation (92 tests) comprising Context Router, Permission Engine, CxRI connectors, reconciliation loop, audit log, and HTTP API (Section 6 ). (4) Five correctness experiments demonstrating zero unauthorized context deliveries, zero permission invariant violations, sub-millisecond freshness detection, acceptable latency overhead, and architectural enforcement of approval isolation—with safety properties formally verified by TLC model-checking across 4.6 million reachable states (Section 7.1 ). (5) Three governance impact experiments on synthetic data demonstrating what goes wrong without governance (phantom content, data leaks, contradictory information) and what the three-tier model catches that alternatives do not (Section 7.2 ). (6) A platform approval survey documenting that no