Minsky's Society of Mind (1986) argued that intelligence is not one thing but a society of small, simple agents — none intelligent alone — whose interaction produces what looks like a mind. Forty years later, LLM agent systems are running the experiment at scale, and the central finding is Minsky's: the magic lives in the organization, not the components.

The modern confirmation comes from both directions. Bottom-up: sub-agents with narrow tool access and focused contexts — a retriever, a coder, a checker — outperform one generalist with every tool and a bloated window, precisely because each society member can be small, fast, and testable. Top-down: frontier-lab systems descriptions converged on layered agency — a model that delegates to specialist processes, each with their own loops — because the alternative, one context holding everything, hits attention economics that no model size has escaped.

Minsky's companion concept deserves equal billing: the B-line of agents — his term for the level that watches lower agents and allocates resources among them. Every production agent society has one whether it is named or not: the orchestrator, the planner, the supervisor. The design question societies keep re-learning is that the B-agents, not the worker agents, determine system character — they decide what gets attention, and attention is the currency everything else is denominated in.

Society-Level Design Rules

• Simple members, rich protocols: capability comes from interaction design; members should be replaceable without re-planning the society.
• Emergence needs checks: societies produce behaviors no member was designed for — which is also how they produce failures no member can explain; run-level observability is non-negotiable.
• Keep the spans shallow: deep agent hierarchies lose signal per layer like any organization; two levels — workers and coordinators — cover most production needs.
• Program the society, not the members: the leverage in prompts, tools, and protocols — member upgrades are commodity; interaction upgrades are moats.

The historical irony is satisfying: the AI field spent decades looking for the intelligent component, and the component era ended with components so capable that their organization — Minsky's actual thesis — became the frontier again.