The model is swappable; the learning stays yours

Microsoft Foundry shipped a piece this morning that’s going to be on a lot of architecture slides by next quarter. “Outcome-driven learning systems: Enterprise RL with OpenEnv and Foundry”, by Govind Kamtamneni, takes the long list of things Microsoft announced at Build 2026 (hosted agents, Toolboxes, Foundry IQ, Memory, Managed Compute, Frontier Tuning, evals, optimizer) and reframes them as one thing: the parts of a learning system you own.

The accompanying post worth reading first is Satya Nadella’s “a frontier ecosystem, not just a frontier model”, which makes the argument explicit. Every firm needs both human capital and “token capital,” and the durable asset is not the model you rent, it is the learning loop you own around your workflows, evals, and accumulated judgment. The line that earned the headline of this note is in the Foundry piece: “the model is swappable; the learning stays yours.”

The two-mode framing is the part most worth sitting with.

A learning system, in this view, improves in two distinct ways.

• Non-parametric learning keeps the model frozen and tunes everything around it: the system prompt, named skills, tool descriptions, the context retrieved through Foundry IQ and Memory, and the model choice itself. No GPUs, no training run, results in minutes. This is what the Agent Optimizer in Foundry Agent Service does for you.
• Parametric learning changes the weights, via Foundry post-training and the ECHO + OpenEnv path. More expensive, slower, and only worth it once non-parametric improvements plateau on your specific outcomes.

The second key idea is that environments and evals are the same artifact. Codify your workflow in a reinforcement-learning environment, attach a rubric that scores “done right” for your outcomes, and you have not just a test; you have a hill-climbing space the agent can practice in and you can grade. That single move (treat the workflow plus the rubric as one object) is the small architectural shift that makes the whole loop work.

Three practical reads for SI and ISV partners reading this in 2026.

• Sell the loop, not the prompt. The deliverable customers will pay for in 2027 is a versioned eval suite, a private RL environment that mirrors a real workflow, a documented optimization cadence, and the wiring that lets the model layer be swapped without losing institutional knowledge. The prompt is a step inside that, not the product.
• Private evals are the new IP. Public benchmarks are a marketing surface. The eval that tells a customer whether their agent is getting better at their actual job is a competitive asset. Help customers build it, then hold it as a versioned artifact alongside the agent code.
• Reach for non-parametric first; reach for parametric only when you’ve earned it. Most teams jump to fine-tuning too early because it sounds like the serious move. The Foundry post is explicit about the order: optimize prompts, skills, tools, and grounding first, and only train weights when the harness is wrung out. That sequencing is going to save customers a lot of GPU and a lot of false starts.

Build’s announcements as a feature list felt scattered. Read as a learning system, they fit together. The partners that internalize this framing now will be a lot more interesting to enterprise buyers in twelve months than the ones still selling integrations.