The rise of the modern data stack
The last decade centralized enterprise data across ingestion, transformation, warehousing, and storage. The promise was that clean data plus SQL would unlock organization-wide analytics.
Enterprise Systems
Your Data Agents Need Context
Recently in the world of data and AI agents, context layers and graphs have become unavoidable topics. That makes sense. Enterprise data agents are nearly useless without the business, system, and organizational context needed to interpret vague questions, reconcile definitions, and reason across messy data environments.
The context on context
It is increasingly difficult to talk to an organization building with data and AI without hearing about context. The market has learned that data and analytics agents cannot reliably answer business questions if they do not understand definitions, entities, source-of-truth systems, and the real operating logic inside the enterprise.
This is not really the model's fault. Enterprises still live on top of partial consolidation, stale semantic files, disconnected apps, and tribal knowledge that never made it into a system of record.
How we got here
The agent frenzy
As LLM capabilities improved in 2024 and 2025, every organization wanted to build chat-with-your-data products, data agents, and AI-driven internal support systems on top of that stack.
Hitting the wall
Those deployments often ran into brittle workflows, weak contextual grounding, and poor alignment with how teams actually make decisions day to day.
The context problem goes beyond text-to-SQL
A question like "What was revenue growth last quarter?" sounds simple. In practice, it depends on how revenue is defined, which quarter convention applies, which data product is trusted, and which business exceptions matter.
Traditional text-to-SQL framing assumes the hard part is getting the model to generate the right query. In many enterprises, the harder problem is that the model cannot know what the user means unless it has access to the context that humans have been carrying informally.
Two common failure points
- Definitions drift. A semantic file may exist, but it can be stale, incomplete, or no longer representative of how the business actually reports today.
- Source-of-truth is conditional. The right table or view can depend on region, time period, product line, or operational nuance that lives outside the warehouse.
Enter the context layer
The core gap is that the agent is not grounded in the business context needed to operate inside the enterprise. That has led to the emergence of the context layer: a maintained, agent-ready layer that helps AI systems understand how an organization's data, definitions, and operating logic actually fit together.
The name varies. Some call it a context OS, a context engine, a contextual data layer, or an ontology. The underlying idea is the same: tie together messy enterprise systems, preserve the business logic on top, and expose that to agents in a usable form.
Why this is more than a semantic layer
Semantic layers are valuable for metric definitions such as revenue, churn, or ARPU. But they are usually hand-authored, tightly coupled to a BI workflow, and narrower than what an autonomous agent needs.
A modern context layer should be a superset. It should capture metrics, but also canonical entities, identity resolution, governance guidance, source selection rules, workflow exceptions, and natural-language instructions that preserve tribal knowledge.
What a modern context layer needs
1. Access the right data
The base layer has to include the warehouse, operational apps, docs, and the unstructured systems where tribal knowledge actually lives.
2. Automate context construction
Usage history, dbt, LookML, semantic definitions, and repeated query patterns can help bootstrap the initial context corpus.
3. Refine with humans
The highest-value context is often conditional, implicit, and historically contingent. Humans still provide the final links that make automation trustworthy.
4. Expose it to agents
Once the context layer exists, agents need real-time access through APIs, MCP, and the workflows where they actually reason.
5. Keep it self-updating
Corrections, new source systems, and changing business rules have to flow back into the context layer so it stays alive instead of turning into stale documentation.
Market direction
Data gravity platforms
Warehouse-centered platforms already own ingestion, storage, and access. They are well positioned to add context-layer capabilities because the data gravity is already there.
AI data analyst companies
Many chat-with-your-data products have discovered that the hard part is not just the interface. It is constructing and maintaining the relevant business context behind the scenes.
Dedicated context-layer companies
A newer category is building the context layer itself from the ground up, ingesting enterprise systems, collecting tribal knowledge, and turning that into an agent-ready substrate.
Looking forward
We are early in the market recognizing the problem and even earlier in building durable solutions. The open questions are fundamental: where should the context layer live, how many copies of it can exist, and what belongs inside the product versus inside the enterprise itself?
That is the thread AgentsNeedContext is following. If you are building around data agents, semantic systems, enterprise ontologies, or context infrastructure, we want to compare notes.
Compare notes