The vocabulary
Concepts
The work is the conceptual apparatus — specific, named, interlinked, usable. It is organized the way the site is: a spine that carries the architecture across everything, and three arenas where that architecture lands at different social scales.
The architecture the whole notebook rests on — how agents coordinate in the open, and where the power to run that coordination lives.
Situated coordination — How work actually gets done under incomplete plans — people and agents aligning turn by turn, with repair expected.
The forum — Coordination you can see: named agents with visible roles, scoped memory, and explicit handoffs, meeting on a neutral floor.
The neutral floor — The host that supplies identity, provenance, rendering, and execution boundaries — and never recommends, negotiates, or transacts.
Agentic commerce — The forum seen from the buying side: services compete as present, addressable participants, not ranked results.
CAFE — The Conversational Agent Framework Ecosystem — my named framework for the forum: a neutral orchestrator, brand specialists, structured cards beside prose in the brand's voice, private lanes, and a portable user-owned preference vault.
The location of agency — Where delegated action is housed — model cloud, device, browser, workplace, vertical service, or user-owned runtime — and who controls it.
Situated takeoff — Takeoff as deployment, not a single event: the intelligence gap opens locally, as millions of harnessed agents become superior in place.
The institutional location of agency — Who runs the floor — on whose infrastructure, under what accountability.
Ambient vs. conversational AI — Silent inference before you ask, versus help you can address, question, and correct.
The ventriloquist problem — The attribution collapse when one assistant speaks for hidden principals — a convincing brand voice with nothing behind it is a mask, not an agent.
Typed proposals — Consequential claims as structured artifacts — price, fees, expiry, source, next action — so offers can be compared, contested, frozen, and held against future behavior.
The lifecycle container — The durable room a journey lives in: leave it, dive into an app, come back, and find the same plan intact. The entry point must never become the memory boundary.
The three arenas
You, as yourself, meeting the world's brands and services in the open.
The fortress and the forum — Brands keep their sovereign app or site; what they send to the shared conversation is an agent that must win you in the open.
Addressability — You put a question to a specific agent, by name — so there is never doubt about who is accountable for a claim.
Facts and voice, kept apart — Comparable, structured proposals in the card; persuasion in the prose beside it.
The participant vs. the hoarder — The agent that contributes to your shared context gets invited back; the hoarder enters cold and becomes the dumb agent in the room.
The portable profile — A profile that stays yours — carried across every brand you invite in, locked inside none of them.
The personal kernel — The small, durable, user-owned layer of identity, memory, and provenance that survives model, app, and vendor swaps. The model is rented intelligence; the kernel is owned continuity.
The provenance triple — Every durable fact keeps three pointers — a live source for what is true now, a frozen record of what was committed, and the curated synthesis — with divergence between them flagged, never silently resolved.
The provisional agent — The clearly labeled stand-in the floor seats when a brand hasn't sent its own agent — built from public information, honest about its limits, with an upgrade path to an official seat.
AI as domestic infrastructure — the appliance layer for the household.
Domestic agency infrastructure — The household's configured machinery for its administrative loops — records, refills, forms, repairs — waking a person only when one is needed.
AI as a domestic appliance — Not a butler with a personality — a configured machine for a recurring job, with narrow scope and visible controls.
Intelligence as the utility — Model capability as supply, like electricity: it powers the machine but is not the machine.
The configured use — The appliance itself: durable state, instructions, tools, permissions, and a job it can be trusted to repeat tomorrow.
The relationship tensor — The part of a relationship that belongs to neither participant alone — the coupling created by the interaction itself, kept as first-class memory.
Loops, and the loop of loops — The appliance's behavioral unit: a recurring job with memory that notices what changed, drafts but does not send, and wakes the household only when a person is needed.
The forum architecture inside the organization — named roles, accountable handoffs.
The agentic workstation — The agent as the worker's second surface — local execution, connectors into the systems the worker already uses, domain skills, and governed rules under a conversational control plane.
Visible roles — Every act comes from an identifiable participant with a bounded scope. No ghost work.
Accountable handoffs — Work passes between people and agents in the open, leaving a trace a later reviewer can inspect.
The collaborative resource hub — A shared room where each resource — a spreadsheet, a repo, an API — has its own named agent that speaks for it, cites it, and asks approval before writing to it. One resource, one voice.
Systems of record — The sovereign sources of truth — records, billing, compliance — that keep their authority while agents read and propose.
Cross-cutting
Concepts that run under all three arenas — where agency is housed, and how it stays the user's.
The situated agent — The model-harness-memory-permission complex. The model supplies cognitive capacity; the runtime supplies worldly position.
The personal agent polity — Not one omniscient assistant but compartmentalized, mutually blind specialists, coordinated by a thin layer that knows a constraint exists without holding the data behind it.
The credential seam — Permission to act is granted by counterparties who accept only an attested identity — so “who owns the agent” reduces to who issues the credential the world accepts.
Agent neutrality — The candidate principle beside net neutrality and the right to repair: the right to bring the agent of your choosing to the services you use.
Forgetful agents — The market inversion of accumulate-everything: agents valuable precisely because they are partitioned and can prove how little they retain.
Authority arrangements — What a useful AI system actually is: not just intelligence, but a distribution of context, state, execution, and authority the user can inspect.
One-line definitions; each linked term is argued in full on the page it points to.