Repository navigation
Where does agent authority sit relative to AG-UI, MCP and A2A? #2790
Replies: 2 comments
|
The tool call about to run feels like the right place, but in MeshKore it ended up out of band. We have no AG-UI integration at all, and no per-tool-call approval or delegated-authority layer. The trust boundary sits at the agent-to-agent request: each call carries an Ed25519 did:key signature proving which agent made it, verified per request at the edge. The evidence trail is signed outcome reports that feed reputation after the call, not a per-action trail next to the run. So of your three options, it's the third, and the AG-UI side just reports what the backend already decided. |
|
Thanks, this is a good question, and the answer is mostly "you don't need anything new". In AG-UI terms, mnki decides and AG-UI carries the conversation with the human. 1. "Acting for this person, with these limits": metadata. 2. Approving the risky ones: interrupts, not a new event. The client can resume on the same thread including custom metadata such as a signature etc. 3. Recording the outcome: attach your IDs to the tool result. So: the decision happens in the backend, the human's part goes over interrupts, and the evidence is carried as metadata. Nothing about it needs a new event type. If several trust layers settle on the same metadata shape, that could be a good basis for a draft proposal later. |
Uh oh!
There was an error while loading. Please reload this page.
I am building mnki, an open trust layer for agents: identity, delegated authority with constraints, and an evidence trail per action. It sits underneath MCP and A2A rather than replacing them, and I would like a protocol designer's view on where it sits relative to AG-UI.
AG-UI standardises the events between an agent backend and the application a person is looking at: sixteen event types, any transport, loose matching. The place I keep landing is the moment a tool call is about to run: the app may want to show that the agent is acting for a specific person with specific limits, ask that person to approve the risky ones, and record the outcome next to the run. Is that an AG-UI event type in your model, an extension, or something the backend does out of band and merely reports?
Twenty minutes would help me; a reply here is fine too. Reference implementation, Apache-2.0: https://github.com/MNKIAgentOS/agent-trust; integration paths: https://mnki.com/docs/integrations. Thank you for AG-UI; a protocol that stays transport-agnostic is the only kind that survives.
All reactions