Autonomous systems are a new security principal
Sentinel treats autonomous agents as first-class identities with accountable owners, declared purpose, tools, data scope, autonomy tier and enforceable runtime controls.
Authoritative agent registry
Production agents should be registered before they obtain credentials, with model, version, tools, data scope, autonomy tier and owner recorded.
Distinct machine identity
Each autonomous runtime should hold a distinct, cryptographically attributable identity rather than sharing a human or generic service credential.
Tool least authority
Tool grants should be treated like privileged entitlements: narrow, justified, reviewable and separated between read and write authority.
Untrusted-content boundary
Retrieved content, messages and external data should remain structurally separated from operator instruction context, especially before sensitive tool execution.
Human approval gates
Irreversible actions should require approval at the tool boundary, with rate limits and cumulative-value caps where appropriate.
Containment
Operators need a single-action mechanism that can revoke an individual agent or halt the fleet without depending on agent cooperation.
Action audit
Agent actions should retain the principal, model version, tool, parameters, result and approval reference needed to reconstruct activity.