The data clouds converged on one bill of materials

An agent that only answers questions needs read access and some idea of what the data means. An agent that does work needs more. It needs somewhere to write what it builds and collects: a submitted form, a reconciled table, the state an automation carries from one run to the next. It needs data and events arriving from other systems. It needs a standard way to call tools, so the same work runs whichever model is driving. It needs an identity with permissions and a record of every call, so somebody can answer for what it did. And it needs the business meaning of the data, or it will compute the wrong number with confidence.

Those five requirements map onto the five parts the platforms bought: an operational store, ingestion, managed MCP, agent governance, and a context layer.

The ledger below lists the moves by company and by part. It is a record of acquisitions and launches in an eighteen-month window, not an inventory of every capability each company has; several already had parts of the design before May 2025. The pattern is in the dates. The operational store and the ingestion pieces landed first, in mid-2025. Managed MCP followed through late 2025. Governance of agent access and the context layer came in 2026.

Two vendors outside the ledger point the same way. Microsoft introduced Fabric IQ with an ontology at its core, a shared vocabulary for agents over Fabric data. Google moved Vertex AI into its Gemini Enterprise Agent Platform, a single place to build, govern, and run agents.

Plotted over time, the convergence is fast. In eighteen months five companies with different starting points, a warehouse, a lakehouse, a CRM, a service desk, and a pipeline business, each added the parts they lacked. None of them added the same parts in the same order, because each started from a different center.

One part changed character during the same window. In December 2025 Anthropic donated the Model Context Protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI and supported by Google, Microsoft, AWS, Cloudflare, and Bloomberg. The July 2026 specification made the core stateless and added MCP Apps and Tasks as extensions, and the main SDKs were downloaded close to half a billion times a month.

That matters for the buyer more than for the vendors. When every model speaks the same protocol, the choice of model and the choice of data platform come apart. A company can pick the model that suits each task and keep its access, its permissions, and its knowledge in one place. Every platform in the ledger now ships an MCP server, so the useful comparison is what each server stands in front of.

Every platform in the ledger assembled the five parts around its own center of gravity. The warehouse vendors built them around data that has been moved into the warehouse. The application vendors built them around the records their applications hold. The pipeline vendor built them around data in motion. Each design works best for the work that already happens at its center, and that is a sound design for a company whose work does.

Most mid-market companies are not organized around one center. Their customer record is in one vendor’s CRM, their orders in another vendor’s ERP, their history in a warehouse, their events arriving from a storefront and a dozen SaaS tools, and their people using whichever assistant they prefer. For them the useful question about the bill of materials is not which vendor has the most complete set of parts. It is whether the parts sit across the systems the work actually touches, and whether they work with every model the company will use.

Plexara delivers the same five parts as one managed deployment that sits across the systems a company already runs. Data stays in the systems that hold it, and queries reach it where it lives. What agents build and collect, from form submissions and webhook events to registered tables and automation state, is kept in the deployment beside those systems. Claude, ChatGPT, Gemini, and the models that replace them reach everything through one MCP endpoint under the same roles and the same audit trail. The catalog and the knowledge graph give every agent the same reviewed meaning across every system.

Deasil runs, upgrades, and monitors the deployment, so the company gets the parts without assembling them. Systems of record stay covers what moves onto those parts one workload at a time, and the capability matrix covers each category in detail.