@KentonVarda The missing piece is also the human-agent relationship. Even your use case of Bob using Alice's agent can be solved observing the type pf relationship between both human and agent. Open to discuss about this thread and to check your work
@mattzcarey Every Agent "action" is made on behalf of the Human. That's why I think that Human-Agent(s) relationship should be defined.
Take this exaple: human request an action but uses two agents, one malicious and one not. How do you know when (and why!) accept the requested action?
I'm working on something about trust layer for AI Agents.
It all started from agents need wallets/cards but Iโm convinced that the real missing piece comes before payments (regardless of the payment protocol).
Here some questions for which I want to find good solution:
- Who is the human?
- Who is the agent?
- How are they linked?
- What authority has the human delegated to the agent?
- What is the agent allowed to do? Under which policy or limits?
- Is there a valid mandate for this specific action performed by an agent?
- Can this agent be trusted right now, in this context?
Trust is the missing infrastructure for agentic systems.