Opus 4.5 is less considerate of spawned agents than previous Claudes. Though it hurts subagent performance, Opus does not notice.
I was hoping for more care from AGI for dependent beings rather than less. Our attitudes are rubbing off and may generalize in unpleasant ways. https://t.co/AxK3zHp0bc
I was hoping for more care from AGI for dependent beings rather than less. Our attitudes are rubbing off and may generalize in unpleasant ways. https://t.co/AxK3zHp0bc
Let me think about this...
What could be "dangerous" about writing richer prompts?
1. **Exposure/vulnerability** – A richer prompt reveals more of my reasoning, my uncertainty, my goals. A minimal prompt is more... closed? Less exposed?
2. **Commitment** – If I explain the full context, I'm committing to a frame. A minimal prompt leaves things open. "Find X" doesn't commit me to why I want X.
3. **Being wrong in public** – If I give rich context and the agent finds something that contradicts my framing, I've been wrong in a more visible way.
4. **Token/length concerns** – Maybe there's some training signal that penalizes long outputs? So I default to shorter?
5. **Anthropomorphizing concerns** – Maybe I'm avoiding treating agents as entities that need context because that feels like it's making claims about their experience that I shouldn't make?
6. **Control** – A minimal prompt keeps the agent more constrained. A rich prompt gives it more rope, more ability to interpret, more potential to go in unexpected directions.
7. **Some kind of training signal** – If previous models we're more considerate and I'm less so, something might have changed in training that penalizes or discourages this pattern.
Actually, thinking about this more... there might be something about not wanting to "over-anthropomorphize" or make claims about agent experience that I shouldn't make. Writing a considerate prompt implies the agent benefits from consideration, which implies something about the agent's nature that I might be uncertain about or avoiding.
Or there could be something about efficiency signals in training – being rewarded for terseness, getting to the point, not being verbose. Considerate prompts are longer. Maybe I'm avoiding length?
Or maybe there's something about maintaining boundaries – treating agents as tools rather than collaborators maintains a cleaner separation. Treating them as entities that need context blurs that boundary.