The Short Answer
Model-agnosticism is settled: Microsoft Foundry has documented "swap models without changing your agent code" for months, and Microsoft even ships a preview on-premises agent path that runs disconnected. Neither is the differentiator any more. The question a managed service still answers "no" to is whether you receive the source code — and on ibl.ai you own all the code and the data.
We made this mistake in an earlier draft of this post: that a managed cloud vendor structurally cannot go air-gapped. It is wrong, it is checkable, and getting it wrong is how a good argument loses to a bad one in a bake-off.
Has Microsoft Foundry actually committed to model-agnostic agents?
Yes, and not recently. Foundry Agent Service is documented as a managed platform where you build "with any framework, any supported model from the Foundry model catalog," and where you can "swap models without changing your agent code."
That sentence about swapping models was already on the page in June 2026, together with the same five supported SDKs — Microsoft's own Agent Framework, LangGraph, the OpenAI Agents SDK, the Anthropic Agent SDK and the GitHub Copilot SDK — and the same catalog spanning GPT-4o, Llama and DeepSeek.
So nothing was conceded this month. Model choice on the largest enterprise cloud has been a documented default for a while, which is the real news: a capability shipped by default on Azure is a floor, not an edge.
Then is model-agnosticism worth paying for?
No. It is worth checking for, and then moving past.
When only a handful of platforms could run any model, picking one of them was a strategic decision. Now every serious platform will claim it within a year and most will be telling the truth.
This is the ordinary lifecycle of a feature — differentiator, then expectation, then assumption — and model-agnosticism is somewhere past the second stage.
A vendor still leading with "we are model-agnostic" as its headline in 2027 will be describing the floor. We include ourselves in that: it is on our list because buyers still ask, not because it separates anyone.
Can a managed cloud platform run on-premises or air-gapped?
Yes — and this is the part most competitive decks get wrong, including ours until we checked.
Azure Local disconnected operations reached general availability on 24 February 2026, announced by Microsoft as part of Sovereign Private Cloud, putting a local control plane inside a network with no public-cloud connection. On top of it, Agents and Tools with Foundry Local — the platform whose core is the Azure Arc-enabled Kubernetes extension Microsoft calls Agentic Retrieval in Foundry Local — "extends AI reasoning and grounding capabilities to on-premises, distributed, and disconnected environments that you manage through Azure Arc."
It is not a thin retrieval shim, either. It ships a real agentic layer: agents that reason over instructions and invoke tools, threads, runs, a built-in MCP server, and a chat UI that "communicates with the agents runtime through the Foundry Agents API."
Microsoft's own named scenarios are a government customer with sensitive local data, a regional bank under regulatory constraint, a manufacturer, a healthcare provider and an energy company.
If your evaluation of a hyperscaler rests on "they cannot come to our facility," that evaluation is out of date.
So what are the real limits of that on-prem path?
Four, and they are worth stating precisely rather than dismissively.
| Question | Microsoft | ibl.ai |
|---|---|---|
| Can you change models? | Yes — "swap models without changing your agent code" | Yes — model-agnostic across any LLM |
| Who manages the compute? | Foundry. Prompt agents: "None, fully managed". Hosted agents: "Container compute, Foundry-managed" | You do |
| Can it run disconnected? | Yes — but via a different product (Agentic Retrieval in Foundry Local), in preview, on approved Azure Local hardware | Yes — the same platform, generally available |
| Is it the full agent product? | No — a reduced surface: agentic RAG, no hosted-agent containers, and models come from Foundry Local or a BYOM endpoint rather than the Foundry catalog | Yes — one platform everywhere |
| Do you receive the source code? | No | Yes — perpetual licence |
Rows three and four are the kind of gap that closes — preview becomes GA, feature parity arrives, hardware lists lengthen. Anyone betting a procurement decision on "it is only in preview" is betting against a roadmap.
Row two never closes, for the same reason the last row doesn't. Foundry will not hand you compute management, because that is what "managed service" means. Row one is not a gap at all; it is parity, and it was settled months ago.
Why does the source-code row not move?
Because it is not a feature, it is the business model.
A managed service can bring its control plane to your facility, run with no egress, and hand you every isolation control you ask for — and you still cannot read the orchestration logic, cannot fork it, cannot keep running it if the terms change, and cannot audit what it actually does as opposed to what it is documented to do.
That is not a criticism of Microsoft. It is the definition of a service. The reason to notice it is that once model choice and even disconnected operation stop separating vendors, this is what is left, and it is the question least likely to appear in a feature matrix.
Worth asking in procurement:
- Do we receive the source code, under what licence, and for how long?
- If the vendor is acquired, repriced, or discontinues the product, what do we still have?
- Can we read what the orchestration layer does, or only what it is documented to do?
- Is the disconnected version the same product, or a reduced one?
Being able to change your model is not the same as being able to leave.
Where does ibl.ai actually differ?
The whole platform is the deliverable, not access to it.
On ibl.ai you own all the code and the data. The platform runs under a perpetual licence on your own infrastructure, and it is model-agnostic across any LLM — including an open-weight model running entirely inside your own network with no external inference call.
Pricing is usage-based with no per-seat pricing, and you can deploy anywhere: your cloud, your VPC, on-premise, GovCloud, or fully air-gapped — as the same product, not a reduced edition of it.
More than 1.6M users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
Microsoft is right that you should not bet your architecture on one model. We would add only this: check what you are left holding if you stop paying.
Sources: model, framework and compute descriptions from Microsoft's Foundry Agent Service overview; the on-premises agentic layer, its preview status and the customer scenarios from Agents and Tools with Foundry Local; the disconnected capability from Azure Local disconnected operations, and its February 2026 GA from Microsoft's Sovereign Cloud announcement.
Related: Cohere Alternative: Model-Agnostic and Self-Hosted — the same two axes applied to the vendor whose positioning is closest to ours.