What is a runtime trust layer for AI agents?

Autonomous systems — LLM agents, workflow bots, MCP servers, humanoids, smart home hubs — all share one dangerous moment: when a planned action becomes a real side effect. A runtime trust layer exists for that moment alone.

Execution is the attack surface

Prompt guardrails filter toxic text. Schema validators check JSON shape. Neither answers: should this email send, this door unlock, this payment execute — right now, from this actor?

Sanctum Runtime sits between reasoning and execution. Every action passes through policy and risk scoring. The runtime returns APPROVE, REQUIRE_VERIFICATION, or BLOCKED— then logs evidence for operators and compliance.

What a runtime trust layer includes

• Interception — one API: verify before execute
• Policy engine — per-action approve / verify / block
• Human-in-the-loop — mobile PWA or console for approvals
• Audit trail — correlation IDs, replay, webhooks
• Local-first risk — Ollama or heuristics when cloud is down

One runtime, twelve categories

The same gate works for agents, ROS2 fleets, smart locks, industrial lines, and healthcare robotics — you define action names; Sanctum enforces policy. See glossary and documentation for integration patterns.

Start building

npm install @sanctum-runtime/sdk

const result = await sanctum.verifyAction({
  actor: 'billing-agent',
  action: 'transfer_funds',
  context: { amount: 4500, currency: 'USD' },
})

Open the console for fleet policy and verification queues, or read Sanctum vs guardrails.