Governance and Safety

Autonomous agents can be unpredictable. Architecture-as-Code allows us to define constraints to ensure safety.

Defining Requirements

Use requirement blocks to specify safety properties.

import { * } from 'sruja.ai/stdlib'


container Agent
container BankAPI

Agent -> BankAPI "Transfers funds"

requirement HumanLoop functional "Transfers > $1000 must require human approval"
requirement PII constraint "No PII should be sent to external LLM providers"

view index {
include *
}

Policy as Code

You can enforce rules about which agents can access which tools.

// Example of a prohibited relationship
// Agent -> ProductionDB "Direct Write"
// ^ This could be flagged by a linter rule

Guardrails

Model your guardrails explicitly as components that intercept messages.

container AgentSystem {
  component UserProxy "Input Guardrail"
  component LLM
  component OutputGuard "Output Validator"

  UserProxy -> LLM "Sanitized Input"
  LLM -> OutputGuard "Raw Output"
  OutputGuard -> UserProxy "Safe Response"
}