When I designed agentic architectures, I started from the agent. I asked what it needed to do its job: which tools, which data, which instructions.
That was not wrong. It was the wrong starting point.
The right starting point is context: what the data means, what the rules are, what the business knows. None of that is specific to the agent. The analyst needs it. The dashboard needs it. The agent is just the newest consumer.
Why it matters more now
Data teams have argued for governance, semantic layers and master data for twenty years. The argument often lost, because a person could always fill the gap.
A finance analyst knows that one legal entity was restructured in March, so the year-over-year comparison is misleading. She knows “manual journal entry” means something slightly different in two systems. Very little of that is written down. It does not have to be, because she interprets the data before anyone acts on it.
An agent cannot do that. It reads what it is given and acts on it. So the knowledge has to be passed explicitly, as definitions, relationships, policies and lineage. Governance and semantics stop being good practice. They become the way we transfer what people used to supply through interpretation.
The agent is the least durable part
You design around what stays. In this architecture, the agent is not it. New models ship every few months, and the agent you build on today will be replaced. Anchoring the design to it means rebuilding when it goes.
What stays is the context and the controls around it: identity, permissions, approval rules, approved tools. Those are what the architecture should be built on, and what the enterprise should own.
Same architecture, different question
Most of the architecture does not change. There is still an orchestrator, still MCP servers, still a tool registry, still retrieval. What changes is how you see them.
Seen from the agent, those pieces are the agent’s equipment, and the design conversation goes to the plumbing. Seen from context, nearly all of them do one job: getting the right context to the agent at the right time. They are pipes. What runs through them is the asset.
Instead of asking which framework, you ask whether the definition of an accrual is correct, who owns it, whether two systems agree on it, and whether you can trace which version an agent read. The goal is the correct slice of context for each request, not more context. Models degrade when the input is full of material that does not bear on the question.
Placement follows from the same view. Semantics live close to the data, in Snowflake or Databricks, next to the queries that use them. The ontology lives close to the agent. It holds the connections between business things, Journal Entry to Account to Cost Center to Legal Entity, and those connections span platforms. It is what lets an agent move from a broad request to the specific records involved.
One request
Take a request: “Accrue $2.4M of September consulting expenses.” The agent works out what it needs. Control checks that the user can create a draft journal. The context layer returns the accrual policy, the legal-entity hierarchy and the account mappings. A decision model like Jev classifies the request as an accrual. The rules engine does the math and confirms debits equal credits. A reasoning LLM steps in only if something is unusual. The journal is created through an approved tool, not by touching the ERP directly.
Three engines, one context. Swap any of them and nothing else changes.
Built for context, not for agents
Because the design starts from context, it is not specific to agents. The same definitions feed a dashboard. The same ontology feeds a decision model in a workflow with no agent at all. The same tools serve an app someone builds next quarter.
An agent platform pays off if the agents succeed. A context layer pays off for every use case that reads from it, including the ones nobody has proposed yet.
Models will keep changing. The meaning of the data stays with the business, and managing it is governance. It is work enterprises should have been doing all along. Most could skip it, because a person was always there to interpret. That person is no longer in the loop.