Authenticated Delegation Between Autonomous AI Agents

This article examines authenticated delegation as a foundational security requirement for autonomous multi-agent systems. It argues that delegating a task is not equivalent to delegating authority: an autonomous agent must be able to verify who granted the authority, on whose behalf it was granted, which actions and resources it covers, how long it remains valid, which constraints apply, whether further delegation is permitted, and whether the authority has been revoked. The article develops a model for verifiable delegation grants, multi-hop authority provenance, monotonic authority narrowing, runtime validation, and graph-aware revocation. It also analyzes replay, substitution, confused-deputy, delegation-forgery, unauthorized re-delegation, and semantic-drift risks. Special attention is given to delegation under intermittent edge connectivity and to Evidence-as-Code as a mechanism for connecting authority roots, delegation paths, policy decisions, protected actions, and execution outcomes. The analysis expresses these requirements as technology-independent architectural invariants within SGAEIA — Secure Governed Autonomous Edge Intelligence Architecture. A practical autonomous incident-response scenario demonstrates how a coordinator, investigation agent, containment agent, and infrastructure tools can collaborate through narrowly scoped, independently verifiable, and revocable authority without relying on transitive trust. This deposit preserves the author-manuscript archival edition of the article originally published on Medium on 11 September 2026. Original publication:https://medium.com/@aridiosilva/authenticated-delegation-between-autonomous-ai-agents-4c1281cd9e14 Archival edition DOI:https://doi.org/10.5281/zenodo.22727194