As LLM Providers Absorb The Middleware Layer, Cognitive Era Data Platforms Face Structural Displacement

The middleware layer that enterprises spent years assembling between their data and their AI is being absorbed by the very LLM providers it was built to connect. Agent frameworks, orchestration runners, and observability platforms that raised hundreds of millions in venture capital are watching their core functions get folded into provider primitives. The implications extend beyond the tooling stack. Cognitive-era data platforms that took over a decade to build are now structurally mismatched for what comes next.

Tom M. Gomez is Managing Partner of Luminity Digital, an applied AI architecture firm that works across decision intelligence, data intelligence, and agentic AI for enterprise clients. He brings more than 25 years of experience across Fortune 500 companies in financial services, healthcare, and manufacturing, with prior leadership roles at Accenture, Capgemini, and Coforge.

As a boutique Applied AI Architect Practice in the US and Europe, Gomez has been writing a multi-part series he calls The Great Compression, which maps how LLM providers are systematically absorbing the layers between model and enterprise. He previously told CIOnews that the next wave of enterprise AI would be decided by who owns the data. This time, he turned to the infrastructure itself.

"If you look at everything architecturally, the answers fall in place. If you look at it as just developing another app, you'll keep writing code," Gomez said. The absorption is already underway. Gomez pointed to how Anthropic's product managers have been signaling the direction for months, telling developers not to build scaffolding because the model would handle it. He said the messaging was deliberate and choreographed, not a one-off product update.

• Primitives over scaffolding: "They put something out and call it a product feature. It is not a product feature. They are defining the infrastructure of the future," Gomez said. "Product managers will gently say, 'Don't build scaffolding. The model will take care of it for you.' What they're really saying is it's all going to be primitives." The implication is that companies building on top of frameworks like LangChain are investing in a layer that providers intend to commoditize.

• The harness problem: The execution harness that gives a stateless LLM the illusion of continuity is where Gomez sees the sharpest compression. Memory, context management, and tool orchestration are all migrating into the provider stack. "The LLM is a stateless machine. Every time you talk to it, it asks, 'Who are you?' The harness is what gives it the illusion of statefulness. And the providers are taking over the harness," he said.

That compression does not only affect middleware vendors. Gomez argued that the cognitive-era data platforms built by companies like Databricks and Snowflake were architecturally incapable of serving the agentic era. Those platforms were designed to ingest unstructured data, process it into lakehouses, feed it to machine learning, and render dashboards for human decision-makers. The human was always in the loop. Agentic AI removes the human, but the deeper break is the substrate itself. Those platforms assumed determinism: fixed schemas, stable identities, repeatable execution. Agentic AI makes nondeterminism the root property, and that single shift breaks coordination, identity, accountability, and security at once.

• Thirteen years versus eighteen months: "The cognitive era substrate had thirteen years to build out. The expansion of the AI market compression happened in eighteen months," Gomez said. "You cannot configure a cognitive substrate to be an agentic substrate. That space is still wide open."

• A missing data layer: Gomez said a new agentic-era data layer would eventually emerge, but its development was blocked by the absence of governance and security standards. The NSA's May 2026 advisory on MCP security reinforced his point. "MCP was simply a USB bus. Everybody assumed it was the new data substrate," he said. "You can't define standards until you define governance and security first. The security protocols need to be in place before you can say yes."

The venture capital dynamics made the situation messier. Gomez noted that investors who backed cognitive-era platforms needed eight to ten years to exit those positions, and many were simultaneously funding the agentic-era companies that threatened to displace them. He pointed to Salesforce's position in Anthropic, built on a 2023 investment, as a telling example. "The very company that is minimizing what Salesforce will be about," he said.

For enterprises watching this play out, Gomez offered a concrete starting point. Companies needed to take inventory of their execution harnesses and understand where the alignment between intent and governance actually lived. Data governance for the agentic era was not the same discipline it had been for the cognitive era, and most organizations had not yet adapted their frameworks to account for the difference.

"You will depend on the provider stack for about 90 percent of what you do. But you still have controls," Gomez said. "You still have time to take control of the architecture. Your data will still be your data. But it is a new data layer."