.png)
Context is the variable that determines AI output quality
We spend a lot of time evaluating AI models on benchmarks, but in practice, model quality is only part of what drives useful output. The more decisive variable is context: what information the model has access to, how relevant that information is to the task, and how efficiently it gets retrieved.
Context engineering is the discipline of getting that right. It covers how you structure prompts, what data you pull in, when you pull it, and how you avoid flooding the model with noise that dilutes the signal.
MCP addresses a core challenge in context engineering. Traditionally, teams would hardcode knowledge into a model through fine-tuning, or pre-load large volumes of data into a prompt and hope the model extracts what it needs. Both approaches are brittle. Fine-tuned knowledge goes stale. Oversized prompts waste tokens and blur focus.
MCP gives agents a better option: retrieve what they need, when they need it, from authoritative sources.
Keeping agents lightweight without sacrificing depth
One of the patterns I find most compelling in MCP adoption is how it changes the relationship between an agent and enterprise knowledge.
Instead of embedding domain knowledge into the model at training or prompt time, the agent uses MCP to pull relevant documents or data at runtime. The agent stays lightweight. It doesn't carry context it doesn't need. But when it does need enterprise-specific information, whether that's a policy, a schema definition, a historical report, or a data asset, it can reach for it on demand.
This has direct implications for analytics and data-intensive workflows. At Incorta, we work with organizations that have deep, complex data estates. The ability for an agent to surface the right slice of that data at the right moment, grounded in actual enterprise context rather than generic reasoning, is what makes agentic workflows viable in production environments.
MCP as the control plane for agent-driven software
Looking ahead, I think MCP's role will expand significantly. Right now, most teams are using MCP servers to retrieve documentation, pull from internal knowledge bases, or connect coding agents to production data. That's valuable, and it's clearly resonating: MCP adoption has accelerated sharply, and context retrieval is consistently cited as its primary use case.
But retrieval is just the starting point.
As software becomes increasingly agent-driven, the demand for agents to not just access context but coordinate across it will grow. An agent working on a data pipeline might need to reason across schema documentation, historical query performance, upstream data quality signals, and downstream consumer requirements, all at once, all in real time.
In that world, MCP becomes the control plane agents use to access context, tools, and actions. It's the infrastructure layer that makes complex, multi-step agentic reasoning practical at enterprise scale. Much like REST APIs became the standard interface for service-to-service communication, MCP-like abstractions are on a path to becoming standard infrastructure for agent-to-data communication.
That's a significant architectural shift. Teams that build with that in mind now will be better positioned as agent-driven software matures.
What this means for data and analytics platforms
The implications are particularly sharp for platforms like Incorta that sit at the intersection of enterprise data and AI.
If agents increasingly rely on MCP to retrieve structured, governed, analytics-ready data at runtime, then the quality and accessibility of that underlying data layer matters more than ever. An agent is only as grounded as the data it can reach. Hallucination in agentic workflows is often less about model failure and more about context failure: the agent didn't have access to the right data, or the data it accessed was stale, incomplete, or unstructured.
Investing in clean, well-governed, high-fidelity data isn't just a data hygiene exercise. In an agentic architecture, it's a prerequisite for reliable AI output.
The takeaway
MCP is already playing a real role in how AI systems retrieve and use context. The teams I see getting the most value from it are the ones treating context engineering as a first-class engineering concern, not an afterthought.
As agent-driven software continues to evolve, MCP will be part of the foundation. Getting ahead of that now, understanding how your data is structured, how it can be retrieved, and how agents will interact with it, is work worth doing today.
Ebrahim Alareqi is the Principal Machine Learning Engineer at Incorta, where he works on AI systems and ML infrastructure. He was recently quoted in InfoWorld's feature on the role of MCP in context engineering.
