Agent Frameworks Are Free Because You Are the Product

Janakiram MSV published an article in The New Stack last week arguing that agent frameworks are the new container wars. AWS has Strands. Google has ADK. Microsoft has its Semantic Kernel Agent Framework. Each one is free. Each one is open source. Each one routes you to a metered runtime where the real money gets made.

He is right about the pattern. He is wrong about what it means.

The Container Playbook, Replayed

The analogy is straightforward and mostly accurate. In 2015, Docker gave away containers. Google gave away Kubernetes. AWS built EKS, Google built GKE, Microsoft built AKS. The orchestrator was free. The compute was not. Every container you orchestrated ran on infrastructure someone was billing you for.

Agent frameworks follow the same logic. AWS Strands funnels you to Bedrock. Google ADK funnels you to Vertex AI. Microsoft’s framework funnels you to Azure AI Foundry. The SDK is the on-ramp. The inference layer is the toll road.

MSV identifies four areas where value is concentrating: context engineering, evaluation and observability, agent security, and interoperability protocols like MCP and A2A. His recommendation: bet on protocols, not frameworks. Pick the interoperability layer as your stable ground.

That recommendation is incomplete.

The Numbers Deserve Scrutiny

Before accepting the framing, examine the evidence. MSV cites several statistics that need qualification.

He claims LangGraph has “80,000+ GitHub stars.” It does not. LangGraph has approximately 25,000 stars. The 127,000-star figure belongs to the parent LangChain repository. Conflating a framework with its parent project inflates adoption metrics by 5x. This matters because the entire argument rests on framework momentum.

He cites Microsoft’s claim that 80% of Fortune 500 companies use its agent tools. Microsoft’s definition counts Copilot Studio and Agent Builder automations within a 28-day activity window. Many of these are low-code workflows, not autonomous agents. The number describes tool adoption, not agent deployment.

He references Gravitee’s finding that 88% of organizations experienced AI-related security incidents. Gravitee sells agent security tools. The survey conflates “confirmed” and “suspected” incidents into one figure. The finding may directionally hold, but the number itself is marketing.

None of this invalidates MSV’s core argument. But when an analyst recommends you “lean into your hyperscaler’s SDK,” it helps to know he is a Microsoft Regional Director, Google Developer Expert, and CNCF Ambassador. He advises hyperscalers professionally. His recommendation benefits his clients. The New Stack, where the article appeared, is owned by Insight Partners, which has invested in CrewAI, Docker, and OpenAI. These affiliations are not disclosed in the article.

Containers Were Homogeneous. Agents Are Not.

Here is where the analogy breaks down, and MSV deserves credit for acknowledging it.

Containers standardized a compute primitive. One container looks like every other container from the orchestrator’s perspective. That uniformity is what made Kubernetes possible. A single scheduler could manage millions of interchangeable units.

Agents are the opposite. They have memory. They hold context windows of varying size. They make decisions that depend on prior decisions. They call tools with real-world side effects. Two agents performing the same task may produce different results based on their context history.

This heterogeneity means the “Kubernetes of agents” probably won’t be a framework at all. MSV argues it will be the protocol layer. Protocols like MCP (for tool interoperability) and A2A (for agent-to-agent communication) could become the shared substrate.

Maybe. But protocols solve interoperability. They do not solve control.

The Real Contest Is Governance

Look at MSV’s four value areas again. Context engineering: deciding what information the model can see. Evaluation and observability: measuring whether agents behave correctly. Security: preventing agents from doing harm. Interoperability: ensuring agents can communicate across boundaries.

Every single one of these is a governance discipline.

Context engineering is access control by another name. Evaluation is audit. Security is policy enforcement. Interoperability without governance is just a faster way for uncontrolled agents to interact with each other.

MSV says: bet on protocols, not frameworks. We say: bet on governance, not frameworks. Protocols enable communication between agents. Governance determines what those agents are allowed to do, what happens when they fail, and who is accountable.

The framework layer is thinning. Everyone agrees on this. The governance layer is thickening. Fewer people are paying attention.

Production Reality Contradicts the Hype

The enthusiasm for agent frameworks runs ahead of deployment reality. Deloitte’s 2026 enterprise AI survey found that 48% of organizations cite orchestration complexity as their primary challenge with AI agents. Fewer than 25% have scaled agent deployments to production.

Gartner projects that more than 40% of AI agent projects will be cancelled or restructured by 2027. Not because the technology fails. Because the governance infrastructure to operate agents safely in production does not exist yet.

Organizations are downloading frameworks. They are spinning up prototypes. They are not running agents in production with the controls those agents require. The distance between a working demo and a governed deployment is where most projects stall.

What to Actually Bet On

The framework choice matters less than MSV implies. If you are on AWS, you will probably use Strands. If you are on Azure, you will probably use Microsoft’s framework. If you are multi-cloud, you will probably use LangGraph or CrewAI. Fine. Pick one.

What matters more:

Context controls. Who decides what information reaches the model? How is that policy enforced? Can it be audited? Context engineering is the new perimeter. Treat it like one.

Evaluation infrastructure. Not whether your agent works in a demo. Whether it works correctly, consistently, under adversarial conditions, at scale. Tools like LangSmith and Langfuse are early attempts. The discipline is nascent.

Security architecture. Given that only 14.4% of organizations have full security approval for their AI agent deployments (per Gravitee, with the caveats noted above), the default state is unapproved agents operating in production. That is not a technology problem. It is a governance failure.

Accountability structures. When an agent makes a decision that costs money, loses data, or violates compliance, who is responsible? The framework vendor? The cloud provider? The team that deployed it? If you cannot answer this question before deployment, you are not ready to deploy.

The Lesson From Containers

The container wars ended with Kubernetes winning the orchestration layer and hyperscalers winning the infrastructure layer. But the real winner was the operational discipline that grew up around both: CI/CD pipelines, GitOps, observability stacks, security scanning, policy-as-code.

The technology was free. The discipline was expensive. The discipline is what made containers production-ready.

Agent frameworks will follow the same path. The SDKs will remain free. The runtime will remain metered. And the governance discipline required to operate agents responsibly will be the actual competitive differentiator. Not which framework you chose. Whether you built the controls to operate agents without losing control of your own systems.

The framework is the on-ramp. Governance is the destination.

Sources: Janakiram MSV, “The reason big tech is giving away AI agent frameworks,” The New Stack (Feb 20, 2026). Deloitte Enterprise AI Survey 2026. Gartner AI Agent Market Forecast 2026. Gravitee State of AI Agent Security 2026 (vendor survey). Microsoft earnings call and documentation. GitHub repository data verified Feb 2026.

Victorino Group helps organizations build governance infrastructure for AI agent deployments. The question is not which framework to pick. The question is whether you have the controls to operate agents in production. contact@victorinollc.com | www.victorinollc.com