Your AI Model Will Change Every 90 Days. Your Stack Shouldn't

If you are evaluating agentic AI right now, the most useful thing you can do is separate two decisions that vendors keep bundling together: which model you use and how your agents connect to your systems. The first is becoming a short-term, reversible choice. The second is a multi-year commitment that determines whether your AI investment compounds or gets rebuilt from scratch every eighteen months.

Two things happened in August 2026 that make this distinction urgent rather than academic.

The plumbing just got standardized

On August 20, Google's A2A protocol — the specification for how autonomous agents discover and communicate with each other — formally joined the Linux Foundation-directed Agentic AI Foundation. That places A2A alongside MCP, Anthropic's protocol for connecting models to tools and data sources, under a single neutral governance body. According to reporting on the announcement, the foundation now counts more than 250 members, including AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.

Strip away the standards-body language and here is what it means commercially. The two dominant ways an agent talks to a tool and talks to another agent are no longer competing proprietary bets owned by rival labs. They are shared, open targets with essentially every major cloud and frontier lab at the table. You can build against them without betting on who wins.

That is a meaningful change in risk profile. Twelve months ago, wiring your ERP, CRM and ticketing system into an agent framework meant accepting a real chance you were building on a format that would be abandoned. Today the interoperability layer is stable enough to treat as infrastructure.

The catch: most vendor agent products still are not built on it. Plenty of platforms in market use proprietary connector formats, and they have every commercial incentive to keep doing so. Which brings us to the second trend.

The model layer is collapsing into a commodity

August was a blur of releases. One public tracker logged twelve new models from seven providers in the first seventeen days of the month; a second counted eighteen confirmed releases from fifteen providers across the full month. The trackers disagree on the exact number, which tells you something in itself — the release cadence is now fast enough that counting it is genuinely hard.

More interesting than the volume is the direction. The efficiency frontier has moved toward agent-capable models small enough to run on a single GPU or a laptop. Nemotron 3.5 Lightning ships as a 30B mixture-of-experts model with roughly 3B active parameters. Muse Glimmer 30B is being pitched explicitly as a local agent that runs continuously. Qwen3.8-27B is out under Apache 2.0. Meanwhile, GPT API pricing has reportedly been cut by around 80%.

Put those together and the conclusion is hard to avoid. Whichever model is "best" for your use case today is a ninety-day decision, not a five-year one. Capability is converging, cost is falling, and deployment options are widening from cloud-only to on-device. Any vendor asking you to commit architecturally to one model provider is asking you to adopt their lock-in as your strategy.

This is why the ROI gap exists

The spending is real. Industry adoption figures for 2026 put 65% of enterprises increasing AI budgets, with a median increase of 22% year over year, and production deployments concentrated in customer service (56%), IT operations (51%) and marketing (48%).

The returns are not keeping pace. The WRITER and Workplace Intelligence 2026 Enterprise AI Adoption survey found that 79% of enterprises report challenges realizing value from AI. More striking, 56% of executives report internal power struggles over AI, and 54% say AI is actively fracturing their organization.

Those are not model-quality numbers. No frontier release fixes an executive turf war or an agent that cannot reach the system of record. They are integration and governance numbers. Organizations bought capability — a model, a license, a copilot seat — and never built the connective tissue that turns capability into a completed business process, or the control plane that lets a risk officer sign off on it.

The durable asset in an agentic deployment is not the model. It is the layer underneath: your tool definitions, your data contracts, your permission model, your handoff logic, your audit trail. That layer encodes how your business actually works. It is also, notably, the part no off-the-shelf vendor is going to build for you. This is the core of what we do in Agentic AI Systems engagements — the model gets chosen last, and it gets chosen so it can be replaced.

The control plane is table stakes now, not a premium tier

A clear shift in enterprise buying behavior through 2026: buyers are demanding audit logs, scoped permission boundaries, kill switches and mandatory human review for high-impact actions before an agent touches a production system. Reporting from the agent ecosystem attributes this to regulatory pressure rather than technical curiosity — which is to say, it is not going away.

If a vendor treats these as an enterprise-tier upsell, that tells you they were bolted on after the fact. Treat governance instrumentation the way you treat logging in any other production system: non-negotiable, present from day one, and cheap to build in versus expensive to retrofit.

We have written before about who carries liability once an agent acts on its own. This post is about the decisions you make earlier — the architecture and procurement choices that determine whether you can answer the liability question at all.

The agent vendor due-diligence checklist

Take these six questions into your next vendor meeting. They are deliberately blunt, and the quality of the answers will tell you more than any demo.

• Is this built on MCP and A2A, or a proprietary connector format? If proprietary, ask what the migration path looks like. If there isn't one, price that in.
• What is the cost and timeline to swap the underlying model? The right answer is measured in days and involves a configuration change, not a re-architecture. Ask them to demonstrate it.
• Who owns the integration code and the tool definitions — us or you? This is the asset. If the vendor owns it, you are renting your own business logic back from them.
• Can you produce an audit trail of every action an agent took last Tuesday? Ask for a real sample export, not a screenshot of a dashboard.
• What is the permission boundary, and what specifically triggers human review? You want a written policy with named thresholds, not "the agent is designed to be careful."
• If you shut down in eighteen months, what survives? Data, integrations, prompt and tool libraries, logs. Get the list in writing.

What to do in the next quarter

Three practical moves:

• Audit your current AI spend for model coupling. For each deployed system, write down how long it would take to switch providers. Anything over a month is a liability, not an investment.
• Fund the integration layer explicitly. Budget it as infrastructure with its own line item, not as an implementation cost buried inside a model subscription. Clean, well-governed connections into your core systems are what make every subsequent model upgrade nearly free. That is the practical value of well-designed LLM integrations.
• Fix the organizational question before the technical one. With 56% of executives reporting internal power struggles over AI, the constraint on your second deployment is likely to be ownership and decision rights, not capability. Name an accountable owner for the agent control plane the same way you would for identity or data governance.

Talk it through with us

Levels AI builds custom agentic systems for companies that want to own their integration layer rather than rent it. We start with your systems, your permissions and your governance requirements, and we choose models last — so that when the next release cycle lands, swapping is a decision, not a project.

If you are mid-evaluation and want a second opinion on a vendor proposal, or you want to map what a portable agent architecture would look like across your stack, get in touch with our team. Bring the checklist. We'll answer all six.