The article focuses on the challenges and strategies involved in securing agents operating in production environments using Palantir's Agentic Runtime. A key aspect is the selection of large language models (LLMs) for agent design. As these models have become more similar in performance, Palantir AIP offers access to various commercial and open-source models through regionally distributed hubs. Setting up model endpoints is straightforward, but ensuring consistent performance under varying workloads is where the complexity lies. Trustworthiness in infrastructure is vital for flexible control loops. Agents can execute based on different triggers, whether from internal updates or external events. The Ontology system provides a unified model for data and actions that both AI and human actors share. Palantir’s Rubix infrastructure supports high availability and scalability, allowing thousands of agent orchestrations to run simultaneously. Each orchestration can have specific fallback mechanisms and retry policies to enhance resilience, with nodes designed to last no longer than 48 hours. Security is tightly integrated into the orchestration of agents. Permissions are determined by multiple factors, including the user who configures the agent and the context in which it operates. This granularity is crucial for production use cases, especially when handling sensitive data like sales opportunities that require geographic restrictions on user access. The article discusses different types of memory—working, episodic, semantic, and procedural—each serving distinct roles in agent operation. The Ontology system supports these memory types while enforcing strict security policies that adapt to each interaction, ensuring that sensitive information remains protected and accessible only to authorized users.