Detailed Analysis
The emerging AI agent economy is being shaped less by the large language model developers and more by a tier of infrastructure companies that control the conditions under which agents actually operate in production. Companies such as Cloudflare, Amazon Web Services, Vercel, Auth0, Okta, Stripe, Datadog, and Microsoft are quietly assembling the layers — runtime, identity, payment, and observability — that determine whether an AI agent can be deployed, trusted, and governed at scale. The article argues that these infrastructure players represent a shift in where AI power actually concentrates: not at the model layer, but at the control layer that governs what an agent can do, who it can act for, what it can spend, and who can stop it.
The runtime control point is the first critical layer the article examines. Because language models are inherently stateless — they process a prompt and return a response without persistent memory — they are insufficient on their own for agentic workloads that require scheduling, recovery from failures, tool coordination, and sustained context across disconnected sessions. Cloudflare addresses this through its Agents SDK, which runs agents on durable objects: stateful microservers with embedded SQL databases, WebSocket connections, and scheduling capabilities. AWS has made a parallel move with Amazon Bedrock Agent Core, bundling runtime, memory, identity, gateway, browser access, and observability into a unified stack. Vercel approaches the same problem through its AI Gateway product, focusing on model routing, budget controls, and load balancing as the primary control surfaces. The convergence across these three platforms on the same underlying thesis — that runtime is itself a governance layer — signals that cloud providers are repositioning themselves as essential intermediaries in any production agent deployment.
The identity and authorization layer represents an equally consequential infrastructure frontier. Traditional identity systems authenticate a human user against application resources, but that model breaks down when an agent acts on behalf of a person across multiple third-party APIs — Google, Slack, GitHub, Salesforce — often with asynchronous approval workflows and access to retrieval-augmented generation pipelines containing documents the user may not be permitted to see. Auth0 has developed a public-facing agent identity framework that addresses delegated authority with explicit constraints: agents call APIs on behalf of users rather than holding broad permanent credentials, token vaults prevent direct secret exposure, sensitive operations require explicit consent, and RAG queries are scoped to documents the acting user is actually authorized to access. Okta, WorkOS, Microsoft Entra, and AWS Agent Core Identity are converging on the same problem space, reflecting a shared recognition that the most dangerous agent in an enterprise environment is not the most capable one — it is the one operating with poorly defined or overly broad authorization.
These developments fit into a broader structural trend in the AI industry: the commoditization of model intelligence and the corresponding rise in value of the surrounding infrastructure stack. As foundation models from Anthropic, OpenAI, Google, and others become increasingly competitive with one another on capability benchmarks, the differentiating factor in enterprise adoption is shifting toward governance, reliability, and interoperability with existing systems. Infrastructure companies like Datadog, which has been building LLM observability tooling, and Stripe, which launched an Agent Commerce Suite enabling agents to execute financial transactions, are inserting themselves at chokepoints that cannot easily be replicated by model developers alone. The practical implication is that the economics of the agent era may ultimately flow less to the companies producing the intelligence and more to the platforms controlling where that intelligence runs, what it knows, what it can spend, and how it is audited — a dynamic that mirrors the history of earlier platform transitions in enterprise software.
Read original article →