Many current AI automation efforts still rely heavily on human intervention to connect different stages, acting as expensive middleware. While individual tasks are automated, the overall workflow remains disconnected, requiring people to manually transfer information between systems. This is inefficient because existing systems like AWS, GitHub, and Jira already have APIs that can share data. The author argues that humans are currently needed to bridge these gaps, but this is a bottleneck that will likely change.The real question is identifying where human involvement is truly necessary, which will vary by team and compliance needs. While AI output isn't always perfect, the author believes humans are often overused in the process. For many tasks, only one or two human decisions are truly required: understanding the initial goal and verifying the final outcome. The author suggests that most of the work between these two points can be handled by machines.A practical example is presented of an automated process for handling a production bug, from error detection to code changes and testing, all without immediate human input. This highlights the potential for significant automation before a human reviewer is needed. The crucial missing piece for more advanced automation is memory, or persistent state, which allows agents to retain and pass on learned information.Currently, this information is often lost in chat windows or terminals, forcing manual summarization. The author proposes that a database or similar mechanism for storing context is essential to enable agents to work together effectively. Building these automated workflows locally on laptops is possible but creates fragile infrastructure that is easily disrupted.These AI-driven automation systems, when performing software development lifecycle tasks, require the same robust infrastructure as any other production service. This includes stability, secure credentials, persistent state, logging, retries, and audit trails. The ultimate goal is for AI systems to autonomously handle tasks like investigating and fixing bugs by morning, eliminating tedious copy-pasting and manual steps for developers.