Research seriesL3paper
A single agent has one identity and one perimeter; a mesh of agents that call, delegate to, and depend on one another has neither. This series builds trust and security models for multi-agent systems from the ground up: why agent meshes break the perimeter and single-identity assumptions that security was built on, how to reason about trust with capability-based, reputation-based, and zero-trust models specialized to meshes, how trust propagates through delegation and where transitive trust turns into confused-deputy attacks at scale, how agents make verifiable claims to one another through attestation and signed provenance, and how it all assembles into a reference architecture of trust brokers, policy meshes, and failure containment. Grounded in NIST SP 800-207 Zero Trust, the NIST AI RMF, the OWASP Agentic Security Initiative and LLM Top 10, MITRE ATLAS, the SPIFFE workload-identity specification, the OAuth 2.0 Token Exchange RFC, object-capability security, and the peer-reviewed literature on reputation systems and zero-knowledge proofs.
One agent has an identity and a boundary; a mesh of agents calling, delegating to, and depending on one another has neither — and the security models built on those two assumptions quietly stop working.
Three trust paradigms compete to govern agent meshes; each answers a different question, and a durable design composes them rather than choosing one.
When agents delegate to agents, authority propagates along a chain — and every hop is a chance to amplify a grant, forget who asked, or lose track of why.
When one agent tells another something, the receiver should be able to check the claim itself — not just trust who is speaking.
The pieces of mesh trust — identity, capability, reputation, verification, delegation — only become a system when a broker, a policy mesh, and containment hold them together.