# Core concepts The five ideas the rest of the docs assume: identity, sponsor, scope, policy graph, and audit log. ## Identity An identity is a distinct, cryptographic name for one agent. It is the unit the directory tracks and the subject of every credential and audit entry. Agents never share identities, so attribution is never a guess. ## Sponsor Every identity is bound to an accountable human sponsor on your team. The sponsor answers for the agent's behaviour and is the approver of last resort when a high-risk action escalates. Money, sensitive data, and new external systems always route to a person. ## Scope A scope is the set of things a credential may touch, expressed narrowly. Mandavo is least-privilege by default: there are no broad standing permissions, and an agent reaches only what its current task requires. ## Credential A credential is a short-lived, task-scoped token issued on demand and rotated automatically. It expires in minutes, so a leak has a small blast radius and a short lifetime. ## Policy graph The policy graph is a machine-readable statement of who may do what, under which conditions. It is the decision point in the hot path: every action is checked against it at execution time, not after the fact. ## Audit log The audit log is the attestation-ready record of every decision. Per-agent attribution with a named sponsor means audit and incident response stop guessing, and compliance evidence generates from the same log you already use.