A control layer built for autonomous agents

Know your agents

• Universal Agent Discovery — process, parent/child lineage, install path, hash, publisher, package; no manual registration per agent
• Agent Passport — every agent carries an identity, not an IP address or a binary name
• Known / Observed / Trusted states — new agents get workable default policies instead of instant bans or instant trust

Decide every action in context

• Capability Model with non-transitivity and time-boxed Capability Leases — revocable, task-scoped grants
• Behavior Chain Analyzer + Process Lineage — sequences are judged, not isolated calls
• Sensitive resource classification and a Security Map that never becomes a treasure map for attackers
• Risk tiers Low → Critical, and a User Hard Deny layer no policy or AI can override

Learn — with brakes

• Adaptive Learning with confidence scores, positive+negative evidence, learning quarantine and trust decay
• Shadow policies and policy replay to test rule changes before they touch live traffic
• False-positive and false-negative feedback loops that never let raw feedback edit security rules directly

Survive mistakes and attacks

• Recovery core: pre-action snapshots, undo, and automatic recovery
• Emergency Lock and Away Mode for one-touch and absence protection
• Anti-tamper: the guarded layer cannot turn off its own guardrails

Stay fast and local

• Fast Path / Slow Path split so ordinary reads never pay for heavyweight analysis
• Local Core First: API expiry, cloud outages, or network drops never stop core protection
• Every security decision can be explained in plain language

Feature availability depends on platform and phase — see Supported Agents and Supported OS for the honest matrix.