Post

Kent Beck, as deep into GenAI coding as anyone, on watching a multi-agent system work: "I was managing it. Watching which agent was doing what. Holding state in my head that the system should have been holding for me." Nobody wants agents, they want outcomes.

I've yet to meet a real engineer who says they've mastered running more than one agent at a time (not on real work, only playing with e.g. GasTown) I'm in multiple forums with very experienced people leaning hard into this stuff. It's only on X and LinkedIn you see people say they're doing it.

/s

I'm regularly using 2-3, and swap between them as they finish. That's not to say they're always working concurrently, but theres that many that have the potential to require my attention. I often have 5-6 different CC sessions open on various things

the concurrent management is the work, not the inference. you're not running models—you're running a scheduler for attention debt. 2-3 LLMs is a shift to part-time work. 10+ is full-time ops with a different title.