A central challenge is closing the gap between what AI agents could theoretically do and what they can reliably accomplish for users. This requires stronger memory, access to relevant information, and the ability to take meaningful actions rather than merely provide advice. The discussion explores the infrastructure needed for an agentic web, where open protocols allow agents to interact with diverse tools, data, and services. Standards such as MCP and NLWeb could reduce organizational silos, but they must be paired with robust security: clear agent identities, permissions, entitlements, monitoring, and risk assessment. Openness and safety can coexist when systems use multiple signals to distinguish legitimate activity from harmful behavior. For software developers, the advice is to remain curious and choose traditional tools, AI agents, or a combination based on the desired outcome. The ecosystem will likely become increasingly specialized, with successful agents defined by how effectively they solve particular problems. As models improve and costs fall, agents will handle more ambitious, asynchronous workflows involving multiple systems and steps, delivering completed results rather than simply answering questions.