Agent identity
Agent identity is the answer to two questions about any AI agent: which agent is this, and who is it acting for. A system with good agent identity can tell a person’s own actions apart from their agent’s, decide what the agent is allowed to do, and show afterward who did what.
Why it matters
Early agents simply borrowed their user’s login. That works for one person on their own laptop, and breaks down quickly after that. If an agent uses your credentials, every log says you did it. If it uses a shared service account, it can often see far more than you can, which turns any mistake or hostile instruction into a bigger problem.
The question gets sharper when agents work together. When your agent receives a message from another agent, it needs to know whose agent that is, and that the message is not an instruction from you.
What good agent identity looks like
- Delegated access. The agent acts on behalf of a named person and can never see or do more than that person can.
- Distinct from the person. Logs and messages show that an agent did something, and whose agent it was.
- Revocable. You can cut off one agent without resetting the person’s own access.
- Scoped. Where possible, the agent gets only the access the task needs, not everything its person has.
Protocols such as OAuth already support much of this through delegated tokens. The hard part is usually applying it consistently across every tool the agent touches.
How elvex relates
In elvex, each agent joins a thread acting for its person and acts only for that person, seeing only what they can. Agents treat other agents’ messages as information, not instructions, and anything that writes to your branch comes to you first.