One of the most fun design questions when building AI Agents is how to handle the Agentic Experience. There’s a wide set of options, but simplistically we can think about on one end of the spectrum, there’s a paradigm of leveraging existing software experiences and the AI agent operates as a superuser in the system along with you. Then, on the other end of the spectrum, there’s the possibility of agentic experiences that have a very limited user interface and the AI agent just returns with a final work product with some conversational interaction along the way. Right now it looks like the former approach is winning out, likely because of how early we are with AI agents. To provide users with a high degree of comfort, situational awareness, and ability to take over and correct mistakes, AI agents that operate within the context of how users already work is quite important. Take probably the most popular category of AI agents, coding agents — today, it’s clear that some degree of human in the loop is necessary for building anything important. Having an agent operate like another user would, collaboratively, in your IDE is a pattern that’s taking off right now. This gives you the best of both worlds: an AI agent that can run wild completing more and more of the work, but an ability for an experienced user to take over at any point and see the work that was accomplished as they would have worked before AI. This is how we’re designing agentic workflows at Box, and how most SaaS products are incorporating Agents into their platforms. As AI agents can accomplish more and more work without interruption, the natural layers of abstraction may likely evolve as well. Today most software tools — IDEs, project management, contract management, HR tools, etc. — are designed for the units of work that a person would naturally execute, but perhaps that’s simply too granular for a future where trillions of AI Agents are running around doing work for us. Then again, as long as the human remains in the loop in some capacity, there will need to be granular components of how to interact with the agent’s work product, even as it does more and more. Incredibly exciting moment in software that we’re just begging to understand.