What makes a good enterprise context layer?
Gartner and Forrester are defining the shape of the enterprise context layer. Across their recent research, six requirements emerge - and they give us a useful way to compare the market.
Marty Pitt@marty_pittContext layer is becoming one of those phrases that suddenly means a lot of different things.
Databricks calls Genie Ontology its “unified context layer”. Atlan calls itself the context layer for AI. Kong has Context Mesh. Confluent has a Real-Time Context Engine.
That’s useful - it means a category is forming.
It’s also a problem, because those products are doing quite different things.
So, what should an enterprise context layer actually do?
Analysts are starting to define the category
Gartner and Forrester have been getting increasingly specific about this.
Gartner’s 2026 research describes a context layer as the architectural plane that gives agents situational awareness, combining semantics, intent, dynamic operational state and runtime delivery. Its agentic integration research extends that into cross-system discovery and action, while its public infrastructure guidance separates probabilistic reasoning from deterministic execution.
Forrester comes at the problem from architecture and ownership, calling for composability, standards-based interoperability, governance, observability and extensibility without lock-in.
Across that research, there are six broad recurring themes:
| Characteristic | What it requires | Research basis |
|---|---|---|
| Cross-system | Span data and capabilities across the enterprise, rather than being confined to one platform or repository. | Gartner describes agents discovering state, reasoning and acting across systems in a real-time context mesh. Forrester calls for standards-based interoperability across enterprise data, metadata and domain systems. |
| Real-time | Reflect the current operational state of the business, not just stored knowledge or historical copies. | Gartner makes real-time dynamic business state part of the context layer, and separately calls for real-time events and synchronisation for autonomous agents in its agentic integration research. |
| Deterministic | Keep trusted planning and execution outside the probabilistic LLM. | Gartner’s public infrastructure guidance calls for an “architectural separation between cognitive reasoning and deterministic execution”. |
| Governed | Apply identity, policy, security, audit and observability across execution. | Gartner calls for delegated identity, auditability and deep observability in its agentic infrastructure guidance. Forrester’s baseline includes step-level logging, policy, authentication, privilege controls and observability. |
| Adaptive | Assemble context around the task, and keep it useful as systems, schemas and semantics change. | Gartner makes intent part of the context layer, filtering enterprise information to task-specific needs. Its Open Semantic Interchange research discusses automated translation, schema change and semantic drift. |
| Open & Sovereign | Keep enterprise context portable, interoperable and under the customer’s control. | Gartner argues for portable, vendor-agnostic semantics. Forrester puts the ownership point more bluntly: “Vendors must fit into their system - not the other way around.” |
Those six give us a useful way to compare a market that’s currently using the same words for fairly different architectures.
So, how does the market stack up?
The comparison below scores the context-layer architecture itself, rather than every adjacent capability a vendor happens to have elsewhere in its platform.
That’s important. Databricks can query external databases with Lakehouse Federation, for example. That’s useful, and it’s why Cross-system isn’t a NO. But Genie Ontology is still primarily constructed from Unity Catalog and Databricks assets, so it lands at PARTIAL, not YES.
The scoring is deliberately simple:
- YES - intrinsic to the context-layer architecture.
- PARTIAL - satisfied only in part, or through an adjacent capability, preconfiguration or replication.
- NO - not satisfied by the context-layer architecture.
- VARIES - depends on the implementation, for protocols or broad categories.
| Vendor | Real-time | Cross-system | Deterministic | Governed | Adaptive | Open & Sovereign |
|---|---|---|---|---|---|---|
| Orbital | ||||||
| Palantir Ontology | ||||||
| Denodo | ||||||
| Kong Context Mesh | ||||||
| Snowflake | ||||||
| Databricks | ||||||
| iPaaS / MuleSoft / Boomi | ||||||
| Atlan | ||||||
| MCP tool layer |
Hover over a rating, or tap one, to see how it was scored.
A few things jump out.
Cross-system access isn’t the same as a cross-system context layer
A platform can have excellent connectors and still be fundamentally centralising.
Denodo is particularly strong here. Its Active Context architecture is built around zero-copy access to distributed operational and analytical sources. Kong attacks the same problem from APIs, while traditional iPaaS products have been connecting systems for years.
By contrast, Databricks and Snowflake can both reach external systems, but their semantic context remains centred on their own platforms. An MCP connector gets you access to another system; it doesn’t automatically make that system part of a shared semantic model.
That’s the distinction behind Cross-system.
Real-time means current state, not fast queries
A context store can answer in 20ms and still be serving yesterday’s data.
Gartner’s context-layer research is clear that dynamic state means live operational state - inventory, order status, customer state - not just historical metrics.
That doesn’t ban caches, indexes or operational replicas. Palantir’s Ontology is materialised, but Palantir describes it as a real-time representation of changing enterprise conditions.
The test is simple: can the agent get the state that’s true now, when it matters?
Dynamic orchestration isn’t deterministic orchestration
Most agent platforms eventually call something deterministic. HTTP, SQL and stored procedures all qualify.
But if the LLM decides which tool to call, in what order, and how the outputs fit together, the orchestration isn’t deterministic.
Databricks defines a deterministic chain as one where the developer predefines the tools and order; its flexible agents let the model choose. Snowflake Cortex Agents use an “LLM-driven plan, use tools, reflect loop”.
So the current market mostly gives you a choice:
- pre-built and deterministic; or
- dynamic and probabilistic.
Doing both at once is the hard bit.
Adaptive means adapting to the task and to change
Gartner’s context-layer framework makes intent part of the layer: the goal filters enterprise information down to what’s relevant to the task.
But the enterprise keeps changing too. Schemas move. APIs evolve. New capabilities appear. Gartner’s Open Semantic Interchange research discusses automated translation, monitoring schemas for change and managing semantic drift.
So Adaptive has two jobs:
Adapt to the goal. Adapt to the world.
That’s different from Real-time. Real-time asks whether the state is current. Adaptive asks whether the layer keeps working as the task or environment changes.
Open & Sovereign is bigger than open storage
If context becomes the primary interface through which AI understands, accesses and acts across the enterprise, the vendor controlling it is sitting in a pretty important spot.
Gartner’s Open Semantic Interchange research argues for portable, vendor-agnostic semantics. Forrester’s “twist-lock” argument says enterprises should define the rails and make vendors fit them, rather than the other way around.
An open table format isn’t an open context layer. Neither is an open-source catalogue, or an open protocol around the edges.
Databricks, for example, has done useful work around open data formats and Unity Catalog. But Genie Ontology - the thing Databricks itself calls its unified context layer - remains a Databricks runtime and representation.
The practical test is:
If we replace the context-layer vendor, can we take the enterprise context with us?
If we keep the tables but have to rebuild the semantics, policies, mappings and operational model, that’s still lock-in.
Where Orbital fits
This is a deliberately narrow comparison.
Databricks is a much broader data and AI platform than Orbital. Palantir has application and operational tooling we don’t. Denodo has mature data virtualisation we’re not trying to recreate.
We’re comparing one thing: the architecture of the context layer.
Our approach is to keep the semantic model in Taxi - an open-source language that can live alongside OpenAPI, Avro, Protobuf, databases and events - and use Orbital as the runtime that plans and executes against those contracts.
The LLM can translate a human request into TaxiQL. After that, it’s out of the execution path. Orbital works out which operations satisfy the request, how their inputs and outputs line up, which policies apply, then executes the plan deterministically.
That gives us:
- context assembled for a new task at runtime;
- live data and actions across existing systems;
- deterministic planning and execution;
- policy and audit outside the model;
- adaptation as systems and schemas change;
- an open semantic model the customer owns.
There are places the bigger platforms are ahead. Palantir’s application tooling is much broader. Databricks has far more mature analytics and ML infrastructure. Atlan has spent years on metadata governance and stewardship.
That’s okay. We’re not trying to become those things.
We’re trying to make the context and execution layer underneath agents open, adaptive and deterministic, without asking enterprises to move everything into another proprietary platform first.
Summary
Context layers are becoming a real architectural category, and Gartner and Forrester are putting useful boundaries around what they need to do.
The six properties are:
Cross-system. Real-time. Deterministic. Governed. Adaptive. Open & Sovereign.
Some are already becoming table stakes. Others - particularly deterministic dynamic composition, adaptation and a genuinely portable context model - are still rare.
If you think we’ve scored something unfairly, come tell us. We’ll update the table when the products change (which, at the current rate, should be sometime next Tuesday).