AI and multi-agent systems
Default framing
Smarter models + more automation = better management; agents as efficient-cause executors of goals.
Juarrero reframing
Intelligence and coordination live in constraint architectures (interfaces, memory, feedback, roles). Automating “management” without understanding constraint structure risks brittle control. Multi-agent systems need designed membranes, enabling protocols, and history/state — not only stronger single-shot pushes.
Three concrete moves
- Specify agent interfaces as membranes (what may pass/transform).
- Make history/state explicit (systems remember — or fail).
- Keep humans in governing loops where closure and values matter; automate catalytic routine carefully.
One experiment this month
(fill: one workflow — redesign constraints around an AI tool rather than only prompting harder)
Linked claims / concepts
Local sources
- Pronovix “shouldn’t automate management yet”
D-poOOaZGT4 - Deliberate complexity talks