What's this about?
The paper argues that agents serving different people or organisations need rules for messages, trust boundaries and accountability. A single-agent tool sandbox cannot manage all multi-party interactions.
What exactly is a social harness?
The protocols and checks around agent-to-agent communication—who may ask whom to do what, and how invalid requests are detected and traced. Think of permissions and receipts for a team of assistants.
What's the community saying?
No substantial 24-hour discussion thread was found; this is an author-reported preprint.
- Cross-principal requests can create new failure modes.
- Message validation and provenance may matter as much as tool access.
- Deployment evidence is needed before treating the framework as solved.