Society of Mind — a Lens on the Fleet
Blue-sky notes, still marinating. This is a lens, not a rename — the fleet's terms stay: ensemble, seat, pipeline, architect, worker, gate, distillation. The point is that Minsky's vocabulary (The Society of Mind, 1986) illuminates the structure the fleet already has — and in a couple of places (agency, K-line) it names something we feel but hadn't named.
The two terms that actually earn their place
Agency — this is the ensemble
Minsky: an agency is a society of simple agents that together do what none can alone; from outside it looks like one capable thing, from inside it's many parts. That is exactly the ensemble — from outside "the reviewer" or "the architect"; inside, N seats + a combiner. The whole fleet is then a society of agencies (review agency, architect agency, worker, gates), and the coding pipeline is an agency-of-agencies.
Where the term does real work: it links straight to dynamic workflows. A dynamic workflow is the act of assembling an ad-hoc agency for one task — fan out these finders, verify with those skeptics, reduce. "Agency" is the noun for the thing a workflow builds. If we ever want a word beyond "ensemble" for a composed, task-specific society, agency is it.
K-line — this is distillation (and the journal)
Minsky: a K-line ("knowledge line") records which agents were active when a problem was solved, so that mental state can be re-activated when a similar problem recurs. Read that against distillation: we freeze a frontier model's judgment into gold sets + tightened exemplars precisely so a cheaper model re-activates that judgment on the next similar case. Distillation is deliberate K-line formation — capturing "what configuration solved this" and making it replayable.
The wave journal is the operational cousin: attempt counts and what-landed reconstitute prior solving-state on replan, so the loop resumes the right mental context instead of starting cold. Gold sets are the durable K-lines; the journal is the working-memory one.
The rest fall out naturally (no term needed, but the resonance is real)
- B-brain watching A-brain — the review/gate layer over the worker — the "green ≠ correct" confirm-vote
- Censors / suppressors — fail-closed gates that veto a bad merge rather than propose a good one
- Difference-engine (means-ends) — replan: goal-state − current-state → the actions that close the gap
- Frames / frame-arrays — the task-spec + metadata schema each leaf is poured into
- Level-bands / heterarchy — epic → feature → leaf; capability at each band, not one flat list
- Papert's principle (growth = new administration, not new facts) — the governance scaffolding (gates, human-merge, audit journal) is the administrative layer that lets more capability be added safely
The spine
Two threads join here:
- Dynamic workflows form agencies — assemble a task-specific society on demand.
- K-lines (distillation) are how the society remembers which agency configuration worked — so next time it reconstitutes the right one cheaply instead of re-deriving it.
That's a coherent account of the whole thing: a society whose agencies are composed on the fly by workflows, governed by censors/B-brains (gates), reasoning by difference (replan), and remembering via K-lines (distillation + journal). Capability accretes inside the agencies; the administrative layer (Papert) is what keeps accretion safe.
Open threads
- Is there a fleet analog to Minsky's "self" — the useful fiction of a single agent? Maybe the unified front door: one CLI/MCP face over a society. Worth poking at.
- Trans-frames (how one frame maps to the next during a change) ↔ the architect's decompose step turning a goal-frame into leaf-frames. Possibly the same move.
- Deliberately don't force the mapping where it doesn't fit. The value is the two terms that click (agency, K-line) and the spine they imply — not a one-to-one dictionary.