Hugging Face team argues against fully autonomous AI agents

A short arXiv paper from four Hugging Face researchers is landing at an awkward moment for the industry's "AI agent" pitch. Margaret Mitchell, Avijit Ghosh, Alexandra Sasha Luccioni and Giada Pistilli argue on arXiv that fully autonomous AI agents should not be developed. Their central claim is that risks to people scale with how much control a user hands to the system.

The paper proposes a five-level taxonomy that is worth borrowing whether or not you buy the conclusion. Level 1 is a simple processor that just prints an LLM's output. Level 2 is a router that picks a branch. Level 3 calls tools the model chose. Level 4 is a multi-step agent that runs loops of its own. Level 5, the one the paper wants off the roadmap, is a system that writes and executes its own code beyond predefined constraints. In their framing, the more control a user cedes to an AI agent, the more risks to people arise.

The risks they catalog are familiar to anyone who has run a real agent past a demo. Cascading errors when statistical models meet real systems. Unsafe actions taken without triggering warnings, from deleting files to unauthorized purchases. Humanlike interfaces that quietly earn more trust than they should. And the tail risk that a broad-access agent gets hijacked into exfiltrating data. Their argued alternative is semi-autonomous systems with human oversight, which they frame as a better risk-benefit profile than full autonomy.

The honest caveat is that this is a position paper, not empirical work. It does not measure incident rates on shipped agents, and it does not name specific vendors, so treat the five-level scale as a governance vocabulary rather than a benchmark. What the paper doesn't give you is any enforcement mechanism, whether the authors want a regulatory line drawn at Level 5 or something softer. Even so, if you sell "agentic" products, expect procurement and safety teams to start asking which level they are buying, and to prefer the answer to be four.