What AI Agents Should Never Do on Their Own

TLDR

Agents are genuinely useful. But give them the wrong task without guardrails and you’ll spend the afternoon cleaning up what took them 30 seconds. This issue covers the specific categories you should never hand off without a human checkpoint, how to write an AGENTS.md that actually works, and the two-agent loop I use for anything medium-complexity or above.

This is the topic I cover in my latest Towards Data Science article, where I walk through the permission categories I gate by default, the AGENTS.md contract I use on every repo, and the two-agent loop that catches what a single pass misses.

Don’t miss it! Read it for free here.