Aaron Levie

Agent Identities Demand New Sandboxing and Access Controls

Agent identities is going to be a super fun and hard problem for software in the coming years. Most agentic systems today assume that the agent can do everything the user can do, and just operate as an extension of that user. This has worked well and is how most auth has worked for cloud software, and it has made integrations super easy over time. But then comes along systems like Openclaw, and suddenly we get a new view into what becomes possible with agents that can operate on their own. And when you have many of them running in parallel. You start to work with them like a colleague, not just as an extension of what you do. But of course this now introduces an all new complexity than the traditional approach. What if you want an agent to access only a small subset of your data? What if you want an agent to have its own sandbox to operate in without any risk of a blast radius if it goes off the rails? What if you want to create an agent that can work with others, without you seeing everything it’s doing? For all of these cases, agents will start to need their own identities inside of platforms. To do this, we likely will need new mental models for how we delegate controls and access to them, how you handle authentication, who gets to manage them in an organization and so on. Lots to figure out in this space right now.