This is a clickable prototype. The agents are simulated and there is no real security. It shows the interaction model.
Ephemerentity is a workspace for temporary agent identities to operate within a bounded context with limited access, permissions, and tasks. The clickable prototype shows how agents can be invoked, work, collaborate, and revoked while maintaining ongoing files and tasks, and keeping data and access partitioned.
Agents are given broad, lasting access to a person's files and accounts, and it is hard to limit what they can reach or to take that access back.
Anyone who gives AI agents access to their files, accounts and tasks has this problem, whether for personal projects or at work. Agents tend to get broad access that lasts, and what they do is mixed in with a person's own work, so it is hard to see who did what or to remove an agent cleanly.
It matters because as people use more agents, one overly broad or forgotten identity can reach data it was never meant to touch.
Instead of giving an agent a lasting identity with broad access, Ephemerentity gives each agent a temporary identity inside a bounded context. The agent has limited access, permissions and tasks. Agents can be invoked, work, collaborate and be revoked, while the files, tasks and history stay with the person. Data and access stay partitioned, so removing an agent doesn't remove the work, and one agent can't reach what it wasn't given.
Prior art on identity and access management of this sort comes from enterprises. These systems are highly technical. A simplified metaphor for access that everyday users can understand is needed. Ephemerentity uses the idea of tabs or groups.
Current access models for agents are heading toward people not trusting the systems. Early problems with consumer agents like Muse come from a gap between what people expect an agent to do and what it actually does. This is an opportunity.
If access is easy to understand and limit, people can match what they expect to what an agent can do. If that holds, people are more likely to trust agents with real work.