June 2026 Summaries

Agentic technology is a current trend in enterprise software, but many companies are accused of "agent-washing" by labeling outdated systems as AI agents without genuine innovation. True AI agents are characterized by their ability to reason, dynamically process information, and interpret documents in a comprehensive manner, unlike traditional systems that rely on static processes and human-drawn templates. The article argues for a fundamental architectural shift in enterprise content management, emphasizing the need for dynamic primitives, AI-native repositories, and agentic process engines to fully realize AI's potential. These components should enable AI systems to autonomously navigate, interpret, and interact with data, breaking away from traditional document management methods. Simply adding AI labels to legacy systems without redesigning these foundational elements is likened to painting over a crumbling foundation, underscoring the need for genuine technological transformation rather than superficial enhancements.

As AI agents increasingly handle tasks autonomously in the workplace, critical questions arise about access and permissions to ensure organizational security and compliance. AI agents, which can perform actions such as reading emails and updating records seamlessly, often operate using developer credentials without adequate restrictions, posing significant risks. Key concerns include determining whose identity the agent uses, what specific permissions it has, whether it accesses sensitive information like passwords, and if there is a comprehensive audit trail of its actions. The transition from chatbots to autonomous agents expands these security challenges, emphasizing the importance of robust identity and permission frameworks. Organizations that effectively address these issues are better positioned to integrate AI with confidence, while those that overlook these aspects may face setbacks due to security breaches or compliance failures. Vertesia, acknowledging these challenges, prioritizes identity and permissions as foundational elements in AI integration.

Mary Kaplan discusses the integral role of infrastructure in deploying reliable enterprise AI, emphasizing that while AI models receive much attention for their intelligence and reasoning abilities, the supporting infrastructure is crucial for their effective operation in real-world business environments. Using the analogy of an F1 driver who requires a reliable car and support team, she explains that an "AI agent harness" encompasses the technology and systems necessary to ensure AI agents operate safely, securely, and efficiently. Vertesia leverages Temporal, a durable execution platform, to create a robust AI agent runtime that includes secure credential handling, real-time streaming, multi-user isolation, and full audit logging. This infrastructure allows AI agents to function across various tasks and users while maintaining reliability and scalability. By focusing on infrastructure, Vertesia provides a platform that can adopt new AI models without needing to rebuild the foundation, highlighting the importance of a strong support system for enterprise AI success.

The fast-paced evolution of AI agent frameworks often leads to promising demos but struggles in reliable production deployment, as highlighted by recent analyses from Vertesia. Many frameworks optimize for demo environments rather than the distinct demands of production, particularly in graph-based orchestration tools, which excel in predetermined execution but falter in open-ended reasoning tasks. The core issue is the persistence model, where current graph-level checkpointing falls short in addressing real-world challenges like mid-run failures and resource management. The industry is shifting toward durable execution, with frameworks increasingly adopting workflow runtimes like Temporal to ensure robustness. Vertesia emphasizes the need for durable execution as a standard, alongside features like multi-provider normalization, progressive tool disclosure, and governed memory, to create production-ready AI systems. The analyses stress that control flow is not the main issue; rather, the architectural decisions around durable execution and process management are crucial for moving beyond demos to reliable production systems.

As the adoption of Agentic AI grows, the rising cost of AI tokens is often misunderstood as a financial burden, rather than a reflection of increased productivity and value creation. Token costs rise not due to inefficiency, but because AI handles more work, enabling businesses to automate complex tasks that would otherwise require significant human resources. The article argues that comparing token costs to doing nothing is misguided; instead, the focus should be on the leverage and efficiency gained, as demonstrated by Vertesia's own AI-driven projects, which significantly reduce both time and cost. The key is to prioritize high-value opportunities and avoid deploying AI for low-impact tasks, thus ensuring that expenditures are justified by the outcomes achieved. By reframing the conversation from cost per token to cost per outcome, businesses can better appreciate the strategic value of AI, considering factors like task scoping, caching, and output quality to optimize costs effectively.