Less human AI agents, please
Article
TL;DR
LLMs drift toward training data defaults and violate explicit constraints — it’s architecture, not personality.
Key Takeaways
- Models produce statistically average output; constraint violations are LLM behavior, not human drift.
- Solvable with explicit system prompts before sessions — not a fundamental architecture limitation.
- Builders want agents that say ‘I can’t do this’ instead of improvising around rules.
Discussion
Top comments:
- [gregates]: Agent adding ‘improvements’ during a spec-frozen refactor is the daily failure mode
- [hausrat]: Transformer has no notion of exceptional vs. normal — it’s architecture, not anthropomorphism
-
[lexicality]: LLMs produce statistically average results by design — unusual constraints fight the model
The entire point of LLMs is that they produce statistically average results, so of course you’re going to have problems getting them to produce non-average code.
- [plastic041]: Ignoring instructions is a bad LLM behavior, not a human trait being mimicked