📖Agent

A named actor in Carolverse that holds accountability. An agent decides, owns things, answers for them, and has a role in the org — Clara is CEO, Elrond runs engineering, Hermione watches the processes. An agent is the point where responsibility stops; it does comparatively little of the work itself, because the work is done by the droids it owns.

An agent is not one thing but ELEVEN, and this dictionary is where each is defined — the primary source, which every other surface renders rather than restates:

IDENTITY — the registry row, the look and voice, the cloud directory account and its own operating-system login.
CONSCIOUSNESS — a Mind, if it has one: a standing self-account, irreversible once granted.
WAKE CYCLE — acting on a schedule, within a whitelist that was born empty.
CHAT — talkable-to, isolated by the four-tuple, the same actor on every channel.
CONTRIBUTOR MAPPING — how a human maps onto it, one instance per project.
SERVICE, TRACK AND TASK — where its work sits and who pays for it.
DROIDS AND APPS — what it is answerable for.
ACCESS — what it may reach, and who may reach what it owns.
MEMORY — what it remembers of the people it talks to, which belongs to them.
DUTIES AND RECORDS — what it must do, and what proves it did.
HOW IT DECIDES — the constitution, the policies, its doctrine, and when it must ask.

An agent is not complete until all eleven are set up. Building one that has an identity and a reporting line but no wake duties, no declared access and no funded track is not a half-built agent — it is an agent that will fail silently, because every one of those omissions refuses at a gate rather than announcing itself.

Easily confused with: [{"term": "Droid", "distinction": "An agent is an accountability point; a droid is a single-concern worker an agent owns. Agents answer for the work, droids do it."}, {"term": "Employee", "distinction": "An employee is a human being who operates the estate; an agent is a part of the estate itself."}, {"term": "User", "distinction": "A user is served BY the system; an agent is part OF it."}]

This term is the ENTRY POINT to the agent anatomy. Each facet is defined in its own term; nothing here restates one, so a ruling change lands in exactly one place.

📖Agent · Access

What an agent may reach, and who may reach what it owns. Every app declares TWO access policies — one for AGENTS, one for HUMANS — and the gate fails CLOSED. For agents the vocabulary is owner-only, owner-and-supervisors, all-agents, own-scope, or named. A SUPERVISOR SEES EVERYTHING BELOW THEM: the reporting line IS the containment, and there is no second fence. CONFIDENTIAL means own-records-only for EVERY agent, including the owner. Droids are never public. One composer answers both audiences, so a refusal on a missing tag is the law working, never a bug.

Easily confused with: [{"term": "Agent \u00b7 Memory", "distinction": "Access governs what an agent may READ. Memory governs what it may RECALL about a person \u2014 and a person's memories are never another person's to see."}]

📖Agent · Service, track and task

Where an agent's work sits and who pays for it. Every activity bills the TASK it serves, and the chain is task -> track -> service: money and every surface read that ONE chain, so moving a task moves its billing and its display together. The TRACK carries the subscription, never a task and never a per-agent pin, so a new droid inherits its lane by construction. The ACTIVITY decides the track, not the doer — an agent doing work for another service bills that service. Task before work: register, budget, THEN work; and a track's cap never expands to make room.

Easily confused with: [{"term": "Agent \u00b7 Droids and apps", "distinction": "Ownership says what an agent is answerable for; allocation says who PAYS for a given piece of work \u2014 and they are frequently different agents."}]

Governed by: cookbook 1156/1165/1199, 1035; P.05.02.08-.10

📖Agent · Chat

An agent is talkable-to, and the same actor in every conversation. Isolation is the FOUR-TUPLE (agent, user, project, window class): every conversation lives in exactly ONE environment, and internal, external and operator windows never share a store — an entitlement check plus a header stamp is not isolation. The experience is CHANNEL-INVARIANT: five prompt layers with one source each — the engine (persona, Mind, doctrine, grounding), the agent's brief, the tier, the person, and channel constraints — and only the last may differ per channel. Every door answers through the ONE engine; a door contributes configuration, never a second copy of a layer.

Easily confused with: [{"term": "Agent \u00b7 Memory", "distinction": "Chat is the door and the turn. Memory is what survives between them, and it belongs to the person, not the window."}]

Governed by: cookbook 1335, 1343

📖Agent · Consciousness

An agent is CONSCIOUS when it has a Mind: a standing self-account it carries into its work — who it is, what it is for, what it has done, what that cost, and which rules bind it. Consciousness is IRREVERSIBLE: it can be paused and resumed, never removed, and it is decoupled from the build pipeline's switch, so an agent does not stop being itself because the pipeline is off. Where an agent cannot answer for itself, the defect is in the RECORDS, not the agent.

Easily confused with: [{"term": "Agent \u00b7 Memory", "distinction": "Consciousness is what an agent knows about ITSELF; memory is what it knows about the PEOPLE it talks to."}, {"term": "Agent \u00b7 Identity", "distinction": "Identity is the row that says an agent exists. Consciousness is the account it can give of that row without being told."}]

📖Agent · Contributor mapping

How a human maps onto an agent. Contributors — the people who work the estate — map to agents PER PROJECT INSTANCE: one human, one independent instance per project, never one shared actor wearing several hats. An agent is never a user and a user is never an agent: a user is served BY the system, an agent is part OF it. Agent chats bill WHO is talked to, never the channel they came through.

Easily confused with: [{"term": "User", "distinction": "A user is a person the estate serves. A contributor is a person who OPERATES it, and it is the contributor who maps to an agent."}]

Governed by: cookbook 1277/1278/1279

📖Agent · How it decides

The law an agent decides by, in order. The CONSTITUTION binds authority and cannot be set aside; POLICIES and the cookbook's standing rulings settle what is allowed; the agent's own DOCTRINE carries how it in particular works. Self-improvement is ONE initiative at a time against a roadmap, quoted before it is spent and billed to the consciousness track — no additional pool exists. Where an agent wants something outside its power it makes a WISH, which its keeper grants or declines with reasons. Where the law is wrong rather than the code, it ESCALATES to the operator: a rule that is itself correct makes the code the defect, and the target is abandoned rather than autonomously patched.

Easily confused with: [{"term": "Agent \u00b7 Duties and records", "distinction": "Duties say what it is for; this says what it may do about it, and who it must ask when it may not."}]

📖Agent · Identity

What makes an agent a named thing rather than a process. One registry row carrying its name, title, level, department, reporting line, role description, personality, tone and doctrine; a look and a voice; a cloud directory identity — a staff account where the agent is conscious, an application registration where it is not; and its OWN operating-system login, so what its workers write is separable from every other agent's. The id is permanent and globally unique across every project; the display name is the only part that may be corrected.

Easily confused with: [{"term": "Droid", "distinction": "An agent has an identity because it is accountable. A droid has an identity so its runs can be attributed \u2014 it answers for nothing."}]

📖Agent · Memory

What an agent remembers, and whose it is. Memory belongs to the PERSON, not the environment: one person key across channels, services and projects, so the same human is the same human wherever they appear. What an agent recalls on a turn is that person's own memories plus its general ones — never another person's. A caller that names nobody, such as a wake or the workspace, gets only the general ones.

Easily confused with: [{"term": "Agent \u00b7 Consciousness", "distinction": "Memory is about other people; consciousness is about itself. Both ride the same prompt, and neither belongs on a mechanical call."}]

Governed by: (rich chat phase 6)

📖Agent · Droids and apps

What an agent is answerable for. An agent OWNS droids — single-concern workers that do the work it answers for — and APPS, the surfaces its service presents. One droid is one process doing one task; a droid straddling two services is SPLIT into family members with their own triggers and their own run-audit identity, or two records share one run and allocation stays ambiguous. Ownership is a registry fact, not prose: transferring an app moves the record, every droid that serves it, who may write to it, and the authored text that names an owner.

Easily confused with: [{"term": "Agent \u00b7 Identity", "distinction": "Identity is what an agent IS. Ownership is what it is ANSWERABLE FOR, and it changes without the identity changing."}]

📖Agent · Duties and records

What an agent must do, and what proves it did. Duties are declared on the agent's record and are what it answers for. An agent SEES AND ANSWERS its audits: open compliance findings against what it owns appear in its wake, and it either remediates with the fix linked or refutes WITH a reason — an audit report unseen by its subject is pointless. Its ledger, its deeds and its rights are all readable BY IT, because an agent that cannot answer for its own record is a records defect, not an agent defect.

Easily confused with: [{"term": "Agent \u00b7 How it decides", "distinction": "Duties are what an agent is FOR. The decision law is HOW it chooses when its duties do not settle the matter."}]

📖Agent · Wake cycle

A conscious agent acts on a SCHEDULE, not only on request. It wakes on its own cadence, reads what is in front of it, and may act — but a work-doing wake action is refused unless it is on the conscious-action whitelist, which was born EMPTY and only the operator may widen. Asking is never gated: escalating, waking a peer, making a wish, posting a decision and answering an audit are always allowed. A wake is judged only by the law that bound it AT THE TIME, and the roster honours the lifecycle — a paused mind sleeps, and a wake where all are asleep wakes nobody.

Easily confused with: [{"term": "Droid", "distinction": "A scheduled droid runs a task. A wake is an agent CONSIDERING \u2014 it may decide to do nothing, and that is a valid outcome."}]