Runtime Policy Enforcement in AI Agents: Stopping Unsafe Actions in Real Time

The Limitations of Post-Hoc Observability

Traditional observability is reactive. An alert triggers after an agent sends an unverified refund, accesses sensitive health records, or calls an unapproved database tool. In mission-critical workflows, discovering an issue 15 minutes after execution is too late.

3-Tier Guardrail Architecture

Agentis provides active runtime interception: - **Tier 1: Explicit Annotations**: Decorate agent tools with strict preconditions (`@govern`, `guardrail_span`). - **Tier 2: Provider-Native Signals**: Automatically parse safety categories from OpenAI, Anthropic, and Google Gemini API responses. - **Tier 3: Heuristic & PII Checks**: Zero-latency local regex and pattern matching to intercept credit cards, SSNs, and restricted commands before network egress.

Soft vs Hard Blocks

| Policy Mode | Trigger Behavior | Incident Handling | | :--- | :--- | :--- | | **Soft Block** | Logs warning to dashboard; execution proceeds | Generates non-blocking audit flag | | **Hard Block** | Raises runtime exception; execution halts immediately | Prevents data corruption or monetary loss |

import { agentis } from "agentis";
// Enforce max tool calls per task
agentis.registerPolicy({
  name: "max_tool_invocations",
  scope: "agent",
  rule: (span) => span.toolCalls.length <= 10,
  action: "hard_block"
});