The extreme usecase of Model Context Protocol: agents access to your full computer, shell, applications and digital accounts to perform tasks in your name. Communication is done through chat, commands are given but the agent performs the task and decides on what steps to ask for human intervention.
This has gigantic security implications: agents have full control over everything including mailboxes, passwords and creditcards and are inherently sensitive to Prompt Injection.
Advanced technically adept users understand that you need to buy separate hardware, telephones and accounts to attempt to sandbox this kind of technology, but the majority will just install the software in search of productivity gains and woefully accept the security risks, and will only regret using it when damages have already happened.
typical agent harnesses
- the original agent harness: OpenClaw learnings feb 2026
- Hermes https://github.com/nousresearch/hermes-agent
- Claude Code
- Codex
- Opencode