The Next Great Open-Source Battle is Over AI Context

For three years, the AI conversation has been a horse race among labs and providers who were focused on developments such as which one of them released the smartest model this quarter? Which benchmark moved? Which context window grew? That race still matters to each group but for the enterprises building on top of them, it matters less every month.

The top models now land within a few points of each other on most business tasks. For example,  prices continue to drop with every release and open-weight models are now trailing the frontier models by months, not years. When the engine becomes a commodity, the advantage moves to whatever the engine runs on and for enterprise AI that fuel is context.

What Context Actually Means & Why it’s Now an Open-Source Fight

Context is everything an agent knows at the moment it acts and is typically comprised of three things. The first is what the agent can access, such as which tables, files, APIs, and documents sit within reach? Next is what the agent understands. Does it know that “revenue” in the finance schema means recognized revenue, while “revenue” in the sales schema means bookings? The third, is what the agent is allowed to do. Can it read customer records? Can it write to them? Can it see salaries, or only headcount?

A brilliant model with poor context gives confident but incorrect answers while a modest model with rich, governed context will give useful responses. When enterprise AI pilots stall, the cause is rarely model quality. The usual culprits are almost always limited access to data, failure to derive meaning from it, and not having the necessary permission to leverage this important information.

Every major platform vendor is seeing the same shift. If context is the new advantage, then whoever owns the context layer owns the customer. That creates a strong pull toward closed context which are proprietary connectors, metadata stores, and permission models that only work inside one ecosystem.

We have seen this movie before. In the data world, storage formats and table formats started proprietary where each warehouse kept data in its own format and readable only by its own engine. Moving to a new tool meant a migration project, so the industry eventually rejected that model. Open source options became far more attractive and evolved into the standard: Apache Parquet for columnar files, Apache Arrow for data in memory, and Apache Iceberg for the table itself. Today, a single copy of data serves many engines. Context is following the same arc only faster. The stakes are higher because context now decides what autonomous software is allowed to touch.

The 4 Layers of Open Context & the Risks of Being Closed

Breaking the context stack into the following four layers, you can see how each one has an open option emerging or is already mature:

●      Tables: Agents need data in a form any engine can query. Open table formats such as Apache Iceberg give agents one consistent view of business data, no matter which engine or tool asks the question.

●      Catalogs: Agents need to discover what exists and who can see it. Open catalog standards such as the Iceberg REST catalog specification and Apache Polaris put discovery and access control in a shared service rather than inside one product.

●      Semantics: Agents need business meaning. A column named amt_net_adj tells a model almost nothing. A governed definition of “net revenue” tells it exactly what to compute. Open semantic standards are the newest layer of this stack, and the most important one to watch over the next year.

●      Agent Connectivity: Agents need a standard way to reach tools and data. The Model Context Protocol (MCP) gives models one interface to many systems. Agent-to-Agent (A2A) gives agents a shared way to talk to each other. Both now live under neutral foundations rather than a single company.

When all four layers are open, an enterprise owns its context which enables the business to swap models, agents, or engines without rebuilding what its AI knows.

Closed context, on the other hand, creates a quieter form of lock-in than closed models. Changing a model is a configuration change and rebuilding years of business definitions, access policies, and metadata is a multi-quarter project.

There is also a safety cost. When permissions live inside each AI tool, every new tool needs its own copy of the rules. But let’s face it, copies drift which means one tool might honor the latest policy while another is running on last quarter’s version. The result is an agent that sees data it should not see and nobody notices until an audit.

When permissions live in a shared open layer, every agent inherits the same rules. You define access once and enforce it everywhere. That design scales, and auditors can easily follow it.

What Leaders Should Do Now

The good news is that organizations do not need to complete a grand architecture before getting started. A few practical moves pay off right away:

1. Inventory Context: List where business data, definitions, and permissions live as most organizations will discover that all three are scattered across a dozen tools.
2. Put Data in Open Tables: Placing every dataset in an open format ensures any future agent can read it without a new pipeline.
3. Centralize access policy in a catalog. Selecting a catalog that speaks an open standard means engines and agents from different vendors will respect the same rules.
4. Write Down Business Meaning: Start with the twenty metrics your leadership argues about most then define them once in a governed semantic layer.
5. Connect Agents Through Open Protocols: Prioritize tools that expose MCP or A2A interfaces over one-off integrations.

The Bottom Line

The first wave of generative AI asked, “Which model should we use?” The next wave is asking, “What does our AI know, and who decides?” Those questions have very different answers depending on whether context lives in open standards or in a vendor’s walled garden.

Models will continue to improve and become less costly, which is good news for everyone. The organizations that win will be the ones that own their context, keep it open, and let any model plug into it.