always-on-agent-safety

What problem does it solve?

An always-on AI agent with episodic memory can accumulate sensitive personal and third-party data, incur runaway API costs, expand scope beyond its value, create unhealthy psychological attachment, and present a high-value attack surface; this guide explains what goes wrong and how to prevent it in practical, actionable terms.

Core Features & Use Cases

• Privacy controls & data hygiene: exclusion lists, retention templates, encryption and export/delete capabilities to minimize unauthorized exposure.
• Cost management: cost-guard patterns, model-cascading, budgeting rules, and alerting to prevent surprise bills.
• Scope and governance: boundary-setting frameworks, human-in-the-loop requirements, and deployment checklists to avoid scope creep.
• Psychological safeguards: language guidelines, agent-free periods, and design patterns to reduce parasocial dependency.
• Security & integrity: input sanitization, memory provenance and hashing, tool sandboxing, monitoring, and an emergency kill switch for rapid containment.
• Use case: evaluate a personal assistant that captures screen, audio, calendar, and messages and produce a deployment-ready safety plan and checklist.

Quick Start

Assess an always-on agent that captures screen, audio, email, and sensor data and produce a prioritized safety plan covering privacy, data hygiene, cost controls, scope limits, psychological safeguards, security mitigations, and an actionable deployment checklist.