For later: people are gravitating to a dead-reckoning model of intelligence by treating length of autonomous runs as a kpi. I think this is ill-posed. Perfectly playable domains allow dead-reckoning to substitute for situated orienting and maneuvering because simulation is 100% fidelity and tractable. Where that’s not true it’s the wrong mental model.

from Venkatesh Rao's Blyg ↗ · snapshot of v1

There’s this ecosystem of terminal-focused agent orchestrators (herdr, conductor, etc) that have this dark-factory view.

I’ve been trying to do sort of the opposite and focus on rubbing as much of my physical senses across what the swarm is building, and force the “dark factory” to orient and gatekeep around that.

I haven’t found a good analogy for it yet. Sort of like “the design team is the ant queen, and the point of the agent colony is to bring things to the ant queen for her decisions” but it’s a tortured metaphor.