Descriptions:
Riley Brown and co-founder An Nandanda break down the AI agent landscape by introducing two core dimensions for comparison: persistence (whether an agent keeps running when your computer closes) and access level (full computer control versus a constrained sandbox). Using that framework, they walk through the major tools on the market — Claude Code, OpenAI Codex, Claude.ai, Manus, OpenClaw, and Perplexity Computer — explaining the design tradeoffs behind each.
The discussion reveals why Claude.ai deliberately limits file system and internet access: to protect non-technical users from irreversible actions. Claude Code, by contrast, runs with full local privileges for developers who need it. Manus sits in the middle as an ephemeral cloud agent that spins up a remote compute environment on demand. OpenClaw, described using Andrej Karpathy’s term ‘claw,’ represents a persistent agent running on a dedicated physical machine — hence the wave of engineers buying out Mac Mini inventory to keep their agents always online.
The second half of the video introduces six practical criteria for choosing the right agent for a task or business context, covering factors like persistence requirements, access scope, reliability, cost, and customizability. For anyone overwhelmed by the expanding field of AI agent tools, this structured mental model offers a reusable way to evaluate new options as the ecosystem continues to evolve.
📺 Source: Riley Brown · Published May 14, 2026
🏷️ Format: Comparison
