The Infrastructure Control Points for Shipping AI Agents

The Control Layer for Production Agents

Production-grade agents require infrastructure that governs execution, identity, and safety. While model performance is critical, the ability to deploy agents depends on five specific control points that manage how agents interact with enterprise systems. Relying solely on the model to handle authorization, state, or safety is insufficient for enterprise environments.

Key Infrastructure Control Points

• Runtime: Agents require stateful environments to handle long-running tasks, scheduled callbacks, and tool failures. Platforms like Cloudflare (via Durable Objects) and AWS (via Bedrock Agent Core) provide the necessary state management and execution context that stateless models lack.
• Identity and Delegated Authority: Agents must operate under constrained, delegated authority rather than broad user credentials. Providers like Okta, Auth0, and WorkOS are developing frameworks to ensure agents only access data and APIs the user is explicitly authorized to see, preventing agents from exceeding their intended scope.
• Governed Data Access: Agents often fail by misinterpreting data or accessing unauthorized context. Platforms like Snowflake (Cortex) and Databricks (Mosaic AI) provide governance perimeters that ensure agents reason only over authoritative, structured, and unstructured data sources.
• Payment and Trust: When agents initiate transactions, they must integrate with institutional trust layers. Stripe is positioning its commerce suite to handle agent-driven issuing, fraud mitigation, and billing, providing the rails for agentic commerce that card networks and traditional payment stacks are currently adapting to support.
• Observability and Kill Switches: Standard logging is inadequate for agent workflows. Teams must implement observability that traces tool calls, cost, and intent (e.g., LangSmith, DataDog, Langfuse). Furthermore, a robust kill switch must be implemented across multiple layers, allowing for intervention at the runtime, identity, or payment gateway level if an agent violates policy or enters an infinite loop.

Operationalizing Agent Workflows

To determine if an agent is ready for production, operators should map every workflow against seven core questions: where the agent runs, who it acts for, what data it can access, what tools it can call, what it can spend, how it is observed, and how it can be stopped. If any of these points lack a clear owner or implementation, the agent poses a significant risk to the organization.