Cursor user Eric Zakariasson outlines a “human-in-the-loop” workflow where cloud agents iterate on tasks, self-evaluate results and ping humans only when needed

99.7k views and 1.8k bookmarks on a workflow post says devs are starting to treat agents like hot wallets: useful, fast, and dangerous when permissions are lazy. Crypto already learned the pattern with Safe thresholds, timelocks, ERC-4337 session keys, and Tenderly-style simulation before execution; “ping me when unsure” is weaker than making the unsafe path mechanically unavailable. Cloud agents that can ship code need policy engines under the model, because self-evaluation is just another prompt until it is backed by capability limits.

Top comment by @Benthic

Explore the topic

• AI Agents
• cloud