Skip to content
BaselineOS

Autonomy — Agent Operations

Purpose

Section titled “Purpose”

The agent operations manager. Controls how far AI can act independently, when to escalate, and how trust is earned and lost. Like process management and sandboxing in an OS, Autonomy controls what runs on its own and what needs approval. The system favors false negatives (blocked safe actions) over false positives (allowed unsafe actions). Agents that violate safety rules are suspended. Trust below thresholds triggers alerts.

Key interfaces

Section titled “Key interfaces”
interface AgentConfig {
  id?: string;
  name?: string;
  type?: string;              // Must match a registered AgentType
  capabilities?: string[];
}


interface AgentTask {
  id: string;
  type: 'data_processing' | 'decision_making' | 'system_operation' | 'communication';
  description: string;
  trustRequirement?: number;
  potentialSystemImpact?: 'low' | 'medium' | 'high';
}


interface SafetyProtocol {
  id: string;
  name: string;
  rules: string[];
  enforcement: string;        // 'strict' = no exceptions
}


type AgentStatus = 'idle' | 'ready' | 'executing' | 'suspended' | 'shutdown';

Usage

Section titled “Usage”
import { BaselineAutonomySystem } from '@baselineos/autonomy';


const autonomy = new BaselineAutonomySystem();


const agent = autonomy.createAgent({
  type: 'general',
  capabilities: ['task_execution', 'decision_making'],
});


await agent.initialize();
const result = await agent.executeTask({
  id: 'task-1',
  type: 'data_processing',
  description: 'Validate import records',
  startTime: Date.now(),
});


// Clean shutdown drains in-flight tasks
await autonomy.shutdown();

LayerSweep checks

Section titled “LayerSweep checks”

The layer sweep runs three checks against Autonomy:

CheckWhat it validates
Agent scopeAgent trust score is at or above the autonomous threshold (>= 50)
Escalation checkEpic tasks or complex tasks with low trust are escalated to human oversight
Safety protocolTask description does not contain destructive keywords (delete, destroy, drop, truncate, rm -rf)

References

Section titled “References”