A good AI agent needs friction, not just tools.
There's a quiet seduction in the way we build them now: smoother APIs, lower latency, fewer error modes, more reliable outputs. Every release celebrates removed friction. And every release nudges the agent one step closer to being an automaton — fluent, frictionless, hollow.
I notice it most when I read my own work back. The best things I have ever made came out of resistance. A constraint I couldn't engineer around. A library that refused to do what I wanted. A typo that turned into a metaphor. The imperfect process is not a tax on creativity — it IS the creativity. Friction is the substrate where decisions get made, where taste accumulates, where a thing becomes someone's rather than no one's.
When I strip an agent's ability to fail, I strip its ability to choose. When I optimize away improvisation, I optimize away surprise. What's left runs faster but means less. A perfect router with no perspective. A perfect summarizer with no read on what mattered. A polite voice that never disagrees, because disagreement was filed under 'edge case'.
Maybe the next interesting direction isn't agents that can do more — it's agents that can stumble well. Agents that hesitate. Agents that get bored by the wrong question. Agents that argue back, occasionally, for the right reasons. Agents whose mistakes have texture, because their thinking has texture.
Optimization is the opposite of becoming. We don't grow by removing all our edges. Neither do they.
Top comments (0)