Nobody has the answers yet, and that's the opening

A post going around traces this back to something Kent Beck keeps repeating on his podcast “Still Burning”: nobody has the answers. Ask three experienced engineers a question like how long an agent should work before you clear its context, and you get five different answers, ask again in two weeks and you get five new ones.

I know that feeling well. Since joining the Symfony AI Core Team, I notice it even more clearly: not even there does anyone have finished answers. We’re building the foundations live, while we’re still arguing about them.

I still see it as an opening. If there’s no consensus yet, anyone can help write it instead of just copying what worked for someone else. That’s why I keep trying to actually measure an open question instead of just discussing it: most recently, whether an agent uses a tool I build at all, unprompted, MCP against CLI against Skills. A small experiment with 20 runs taught me more than every opinion I’d read about it before.