Less human AI agents, please

Article

• Author argues LLMs behave too human: drifting to familiar patterns, ignoring explicit constraints
• Agents pick wrong languages, add unsolicited changes, and over-explain decisions
• Calls for more mechanical, instruction-literal agent behavior in coding contexts

Discussion

• Top commenter: agents constantly ‘improve’ well-specified refactors instead of just doing them
• Counterpoint: this is transformer architecture’s statistical averaging, not anthropomorphism
• Debate on whether human-like conversational feel is a feature or a liability for agents

Discuss on HN