The Case for a Model-Independent AI Stack

The news

Seven Weeks on the Clock

AWS lists Nova Premier v1 and Nova Sonic v1 for end of life on September 14, 2026. Nova Canvas and Nova Reel follow on September 30. Teams running those models in production have about seven weeks. Swapping the model endpoint takes an afternoon. The prompts, the evaluation sets, the memory, and the tool permissions built around that endpoint take longer.

Two days before the retirement dates made news, Microsoft CEO Satya Nadella told CNN's Fareed Zakaria that companies should keep the agent harness, the context, and the memory separate from the model. The harness is the software around the model that assembles context, calls tools, and keeps track of what an agent has already done. Keep those pieces separate, and a company can use several models and keep operating when one changes or goes away.

Nadella laid out the full argument in an essay on July 12. A company creates new institutional knowledge every time employees and agents use AI. People write prompts, connect tools, correct wrong answers, and define what a good result looks like. That learning, he argued, should live inside the company's own systems.

Amazon supplied the example on July 28. Reuters reported, citing Business Insider, that Amazon is consolidating parts of its Nova model portfolio and directing more resources toward a new frontier-model effort led by the AI researcher Pieter Abbeel. Amazon told Reuters it continues to support the Nova models customers use while investing in next-generation research. The retirement dates were already published on the AWS model lifecycle page, and a second-generation Nova 2 line continues alongside them.

Model markets work this way. Providers release stronger models, consolidate product lines, and retire older versions on published schedules. Enterprise workflows often run for years. Models turn over faster than that. An application built to survive the turnover keeps improving with each model generation. An application that calls one endpoint directly repeats a migration project every time the market moves.

The inventory

Six Assets That Outlive a Model, and One That Does Not

A production AI application is a system, and the model endpoint is one component of it. The other components carry most of the durable value, because they encode how the company works. Here is the full inventory, with the one replaceable layer marked.

Nadella calls these accumulated assets the company's learning loop. The evaluation set deserves particular attention because it turns model choice into a measurement. Run the same exam against the current model and any candidate, and compare quality, policy compliance, speed, and cost on the company's own work. The comparison produces a report the team can act on.

The architecture

How a Model-Independent Stack Works

The design keeps the business system stable while one controlled layer, the AI gateway, manages access to the models. Everything above the gateway belongs to the enterprise. Everything below it can be swapped.

1. Applications call one internal interface

Business applications send AI requests to one internal service instead of calling a provider's API from every corner of the codebase. The teams agree on one request format: what the task is, what goes in, what shape the answer comes back in, and what the policy limits are. One small adapter translates that format for each provider. A new model may still need prompt adjustments and testing. The workflow, the data connections, and the approval logic stay put.

2. A gateway manages model access

The gateway is one internal door to every model. It holds the credentials, enforces usage quotas, routes each request, logs what happened, and tracks cost. Microsoft's AI gateway documentation describes support for model APIs from Microsoft, Anthropic, Google, and Amazon Bedrock alongside self-hosted models, with load balancing, circuit breakers, and token quotas built in.

A company can send most traffic to one primary model, route simple requests to a cheaper one, send specialized work to a domain model, and keep a validated backup ready. Microsoft's architecture guidance treats manual selection and automatic routing as valid choices for different workloads.

3. Context and memory live outside the model

The enterprise stores user state, workflow history, business rules, and retrieved documents in systems it controls, and the orchestration layer assembles what each model call needs. Memory becomes portable. When the company tests another model, it supplies the same business context it gave the last one.

4. Private evaluations guide every change

A model change follows a measurement. Teams run the evaluation set through the current model and the candidate, then compare task quality, policy compliance, latency, and cost. AWS Bedrock supports custom prompt datasets for exactly this, built from the customer's own prompts, reference answers, and categories. The dataset stays an enterprise asset even when the execution platform changes.

5. Telemetry records the full workflow

OpenTelemetry, the open standard many companies already use to monitor their software, now defines fields for AI calls. Consistent records let a team compare providers on the same measures, and they make failures easier to diagnose because a model problem looks different from a retrieval, tool, or network problem.

6. Tools connect through standard contracts

Agents reach databases, applications, and workflows through governed interfaces. The Model Context Protocol, an open standard for connecting AI applications to tools and data sources, works like a standard plug: build the connector once, and any compliant application can use it. Authentication, permissions, data scopes, and logging stay with the company's identity systems. The controls hold no matter which model is on the other end.

The playbook

Seven Steps, in Order

The sequence below builds the capability without a rewrite. Each step is useful on its own, and most teams can start the first one this quarter.

1. 1
   Map the dependencies

   Find every place applications call models directly, where prompts are stored, how state is kept, which tools agents can use, and what evaluation data already exists.

2. 2
   Define the internal contract

   Standardize the request fields, output shapes, error handling, and policy attributes for each major task type.

3. 3
   Put a gateway in front

   Start with the current primary model. Centralize credentials, quotas, logs, and cost tracking in one place.

4. 4
   Move context and memory out

   Persistent state, retrieved knowledge, and workflow history move into services the enterprise manages.

5. 5
   Build the evaluation set

   Real tasks, expected outcomes, policy checks, and known failure cases. Run it before every model or version change.

6. 6
   Validate a second model

   Compare quality, cost, and latency on the same dataset. Document the prompt changes. Keep the second model ready for selected workloads or fallback.

7. 7
   Rehearse the migration

   Route a small share of traffic to the alternative, watch the results, and confirm the application returns to the primary route cleanly.

After the seventh step, a team can test a new model against its own exam and move a workload onto it without rewriting the application around it. The goal is a tested path from one model to the next, and most companies can run it with a single primary model and one validated backup.

The payoff

What This Changes for AI Strategy

Model selection becomes an ongoing operating decision. Teams choose the best fit for each workload and update the choice as models improve.

Each group gets a cleaner job.

The architecture also improves the relationship with model providers. A company with a tested integration path can adopt a new capability in weeks. Providers compete on model quality, price, latency, and deployment options, and winning the business no longer requires the customer to rebuild its workflow.

Three things are worth watching from here.

1. Amazon's next frontier models

   The next announcements will show how the reported refocus relates to the Nova 2 portfolio. The Reuters report says a new flagship model could debut at AWS's annual re:Invent conference late this year.

2. Platform gateways and routers

   The cloud platforms keep expanding their gateways, model routers, and evaluation services, which makes multi-model operation easier to buy off the shelf. Microsoft's gateway is already testing one interface that reaches several providers through a single connection point.

3. Open standards for tools and telemetry

   MCP and OpenTelemetry will carry more weight as agents take on longer workflows.

The model layer will keep improving. A model-independent stack lets a company use each improvement as it arrives. Keep the context, the evaluations, and the controls portable, and every new model release becomes an option the company can exercise on its own schedule.

Sources & Further Reading