Everyone Has the Same AI Models. Here's What Sets Agents Apart with Richmond Alake, Oracle

The industry is moving beyond prompt, context, and control planes toward an "intelligence plane" - an architectural layer that gives agents knowledge, memory

Beyond the Control Plane: Toward an Intelligence Plane for Agentic Systems

Abstract

This analysis examines a proposed architectural layer for agentic systems termed the intelligence plane, positioned as the successor to the prompt layer, context layer, and the currently dominant control plane. The central thesis is that parity in model access and information access does not confer intelligence or competitive differentiation; differentiation instead arises from owned memory, governed data access, semantic structure, and continual learning. The argument is developed through a historical analogy to World War II cryptanalysis, a reference architecture mapping control-plane components to intelligence-plane extensions, and two demonstrations. Reported results indicate a governed analyst agent achieving correct answers 99% of the time using 32,000 tokens versus 200,000 tokens for a naive baseline, with a semantic layer reducing a definitional error margin to 5%. Practical implications include treating memory as governed data, capturing interaction traces as a proprietary asset, and fine-tuning small decision models for memory routing rather than hand-written logic.

1. Introduction

Agentic system design has progressed through identifiable tooling generations. Early practice concentrated on the prompt layer, optimizing single-turn instructions. Attention then shifted to the context layer, concerned with assembling relevant information within a bounded window. The present generation is characterized by the control plane: a centralized management system governing agent identity, policy and guardrails, orchestration, permissions, and auditing.

A parallel vocabulary has emerged around the software factory, a term coined by Bob Bemer at General Electric in 1968 to describe measurable, repeatable program creation. The term has been repurposed to describe agentic systems that produce software through multi-agent parallel execution, reflecting an industry preoccupation with throughput and orchestration scale.

The thesis advanced here is that these layers, while necessary, are insufficient. Orchestration and governance determine what an agent is permitted to do; they do not determine what it knows or how it improves. The proposed intelligence plane cuts vertically through the agent stack, encapsulating four concerns: what the agent should know (knowledge), what it should remember (memory), what it is allowed to see (governance and data access), and how it gets better (continual learning). The organizing claim is stated compactly: "information does not equal intelligence." Section 2 establishes memory and plane-separation foundations. Section 3 analyzes the architecture and supporting evidence. Section 4 extracts implementation findings. Sections 5 and 6 discuss implications and conclusions.

2. Background and Related Work

Agent memory is defined as the mechanism by which agents retain, reuse, recall, and refine information - the "four Rs." The accompanying taxonomy distinguishes temporal scope (short-term versus long-term) from cognitive function, borrowing from cognitive psychology: semantic memory (facts and definitions), episodic memory (specific past interactions), procedural memory (learned routines), and working memory (active task state). This framing, characterized as a fringe position roughly three years prior, is presented as now broadly accepted - an adoption trajectory used to motivate the plausibility of the newer intelligence-plane construct.

Plane separation itself is borrowed from networking literature, which has proposed a knowledge plane layered above the management and control planes. The intelligence plane is offered as the agentic analogue: control planes route and enforce, whereas knowledge planes reason over accumulated state. This framework is summarized as a composition of equations: Agent = Model + Harness; Control = Governance + Security; Intelligence = Knowledge + Memory + Learning. The third term is characterized as producing a multiplicative rather than additive effect on overall system capability, implying that knowledge, memory, and learning interact rather than simply accumulate.

3. Core Analysis

3.1 The AF-is-Midway Analogy and the Information/Intelligence Distinction

The argument for an intelligence plane rests on a historical case drawn from World War II codebreaking. Two U.S. groups - a centralized unit in Washington and a smaller unit, Station Hypo, in Pearl Harbor - received the same intercepted Japanese communications referencing a code word, "AF." Washington possessed more centralized resources and control, yet Station Hypo, operating from a basement and recording intercepted messages on paper over months, correctly deciphered "AF" as Midway Island, while Washington reached a different conclusion from identical raw data. The explanatory variable was not access to information but the accumulated, retained record functioning as memory, which enabled correct inference. This analogy is used to ground the claim that equal access to models or data does not produce equal outcomes; the differentiator is what is retained, structured, and reasoned over.

3.2 Reference Architecture: Mapping Control Plane to Intelligence Plane

The proposed reference architecture maps existing control-plane components - gateway/connectivity, orchestration, memory core, and infrastructure/platform - to intelligence-plane additions: identity-linked data access, context assembly, ontology/semantic layers, and inference/fine-tuning/continuous learning. This mapping suggests the intelligence plane is not a replacement for control-plane infrastructure but an overlay that adds knowledge structure (ontology, semantics) and learning capability (fine-tuning, continuous learning) onto existing orchestration and governance mechanisms. The architecture's viability is attributed partly to maturation of the surrounding ecosystem: frontier model scaling laws have made model capability more predictable, open-weight models now trail three to six months behind frontier releases, and decision models - smaller models used at specific pipeline decision points - can now be embedded throughout retrieval pipelines rather than only at the final generation step.

3.3 Empirical Demonstration: Governed versus Naive Analyst Agents

The architecture's practical effect is illustrated through a comparison of a "governed" analyst chat agent, built on the full intelligence-plane reference architecture including ontology and semantic layers, against a "naive" analyst agent lacking these layers. The governed agent achieved correct answers 99% of the time using approximately 32,000 tokens, compared to 200,000 tokens consumed by the naive agent. Two specific failure modes in the naive agent were attributed to missing structure: an incorrect resolution of the entity "counterparty" (attributed to reliance on keyword search rather than an ontology layer) and an incorrect definition of "concentration" (attributed to the absence of a semantic layer, with the governed agent reducing the resulting error margin to 5%). In finance and other mission-critical domains, this magnitude of error is framed as consequential, potentially separating regulatory compliance from material fines.

4. Technical Insights

Several implementation-level findings follow from the demonstrations and are presented as low-effort starting points for building toward an intelligence plane:

• Memory routing - classifying and directing model or context outputs to the appropriate memory store (for example, conversation memory versus workflow memory) - is identified as a high-leverage, low-effort entry point. A fine-tuned decision model based on Qwen, trained on classified conversation data, was used to perform this routing in place of hand-written routing logic, reported to improve quality and in some cases reduce latency relative to rule-based approaches.
• Three model types are described as embeddable within an agent memory system: a decision model (for routing and classification), an embedding model, and a re-ranker. This modularity allows memory infrastructure to incorporate model-based reasoning at multiple points rather than relying solely on a single large generation model.
• Governance propagation is treated as a data-layer concern rather than an agent-logic concern: end-user row-level data access permissions are propagated into the agent's identity so that the agent inherits only the data access the acting user is permitted, rather than operating with broader standing privileges.
• Deployment flexibility is addressed through infrastructure capable of hosting embedding models, inference, and re-rankers in air-gapped or on-premises environments, alongside hosted cloud inference options, indicating that intelligence-plane components are not presumed to require a single deployment topology.
• A key trade-off noted is token efficiency versus architectural investment: the governed agent's 6x token reduction (32K vs. 200K) was achieved only after investing in ontology and semantic layer construction, suggesting upfront structuring cost is recovered through reduced inference cost and improved accuracy over time.

5. Discussion

The central implication is a reframing of competitive advantage in agentic AI away from model selection and toward owned, structured memory and governed data access. Given that frontier models are increasingly commoditized and open-weight alternatives trail by only months, differentiation premised on model choice alone is argued to be unsustainable. This parallels broader industry discourse around retrieval-augmented generation and knowledge graphs, but extends it by framing memory itself - not merely retrieval - as the proprietary asset, and by proposing that memory engineering, data governance, and model fine-tuning be treated as a unified architectural layer rather than disconnected concerns.

A notable gap is the reliance on a single vendor's demonstrations (governed versus naive analyst agents) as the primary quantitative evidence; the 99% accuracy and 32K-versus-200K token figures are not independently benchmarked or compared against alternative architectures such as standard RAG with reranking alone. Future investigation would benefit from controlled ablations isolating the marginal contribution of the ontology layer versus the semantic layer versus the decision-model routing, rather than evaluating the full intelligence plane as a bundle.

This direction connects to a broader industry trend of embedding smaller, task-specific models throughout pipelines rather than relying on a single large model for all steps, and to increasing attention on continual learning occurring across token space, weight space, and latent space (for example, via prompt caching) as a lower-cost alternative to full retraining.

6. Conclusion

The intelligence plane is presented as a necessary architectural addition to existing prompt, context, and control layers, addressing what agents know and how they improve rather than merely how they are orchestrated and governed. The core contribution is a reframing of competitive differentiation: given commoditized model access, advantage accrues to systems that own structured memory, enforce governed data access, and support continual learning. Practically, this suggests engineering teams prioritize memory-as-data practices, capture interaction traces systematically, and introduce fine-tuned decision models for routing tasks as an accessible first step before pursuing fuller ontology and semantic layer investments.

Sources

• Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub