Essays, field notes, and practical frameworks for identity systems that have to work in the real world.
A practical OAuth architecture for MCP separates client, authorization server, protected resource, and audience-bound access token.
An agent should propose tool calls, not authorize them. Every execution needs schema validation, policy enforcement, scoped credentials, and an audit trail.
An MCP server is a protected resource, not an OAuth client wearing a different name. It must publish trust metadata and validate audience-bound access tokens.
MFA raises confidence at login, but modern applications still need secure sessions, risk checks, authorization, and revocation after authentication.
Enterprise SSO is a federation of trust, not one shared login session. SAML or OIDC carries identity proof, while each application still owns its session and authorization.
A secure authentication API is a protocol boundary, not a login controller. Token semantics, failure behavior, rotation, and agent identity must be designed explicitly.
Sessions are distributed security state, not just cookies. At scale, rotation, revocation, tenant context, and agent delegation must survive every service hop.
Multi-tenancy is an authorization boundary, not just a tenant_id column. Identity must establish tenant context before every access decision.
RBAC works for stable job functions. AI agents often need ABAC's contextual, per-request authorization decisions.
Identity is not a security review item you bolt on before launch; for AI applications, it is architecture that needs to run through the whole system from the first commit.
Authentication answers who you are. Authorization decides what you can do, and AI agents need that decision checked continuously.
Four identity standards, four different jobs. A practical guide to understanding how OAuth, OIDC, SAML, and JWT fit together in modern applications and AI agent systems.
API keys prove that a caller knows a secret. AI agents need scoped, expiring, and auditable identity instead.
OAuth’s human-centered assumptions start to break when agents act unattended, delegate work, and chain decisions across services.
Enterprise identity solved many of the problems AI agents are creating today. Here are five lessons worth borrowing.