Glossary — who is who

One word for each thing. These docs, the engine’s code and the dashboard use them the same way.

The system

• Operator — the whole system: your brain (how your company works) plus the engine that runs it. Never an agent, never a person. “What can my operator do without asking me?” asks about the system.
• Brain — your repository: config, features, prompts, memory.
• Engine — the framework the brain runs on.

Agents

owner(s) ──► chief ──► lead of feature A ──► agents
                  ├──► lead of feature B ──► agents
                  └──► agents (a feature without a lead)
outside the chain: other bots' personas · boss · system agents

• Chief — the role that holds the whole operation: it knows every feature and delegates inside it. It is the agent the owners talk to by default: an owner’s DM with the chief’s bot, the dashboard chat, the daily brief, “Fix with agent”.
  • Configured by defineProject({ chief: { name, role } }). name is its persona name (engine default chief), role the role its chats run (engine default system:chief, file constitution/Agents/system/role-chief.md).
  • The UI shows the persona name (omni-knight, largo), never the word “chief”.
  • It is a role, not one row: each owner has their own DM with it — one role, one memory, separate conversations. The dashboard chat shows the primary owner’s DM.
  • The bot in bots[] that omits role is the chief’s bot; at most one bot may omit it.
• Lead — an optional agent that looks after one feature, so the chief hands that feature’s work to it instead of to individual agents. A role plus a heartbeat, by convention: features/<feature>/<feature>-lead.agent.md and <feature>-lead.heartbeat.md — each tick a fresh agent reconciles the feature against its map (FEATURE.md) and escalates to the chief. The build-an-operation engine skill writes them; brain-example’s freelancer-lead is the worked example.
• Agent — any running instance of a role. The ones the chief or a lead delegate to are just “agents”.
• Harness subagent — a helper Claude Code spawns inside one turn with its own Agent tool (capability harness:subagent). Not an agent in this glossary’s sense: no row, no role, model or grant of its own, gone when the turn ends. It works with the agent’s tools but not under its role: only the prompt the agent writes for it. Every other “agent” or “sub-agent” in these docs (agent_task, agent_create, capability agent:execute) is an engine agent, with a row of its own.
• Chat agent — the mechanism behind a chat: one agent per (chat, bot), created on the first message and named after the bot’s persona. Which role it runs is the chat’s route (chat-access.md): the chief in an owner’s DM with the chief’s bot, the counterpart’s secretary role in a business chat, the rule’s role in a group. A thread agent is the same per discussion thread or forum topic.
• Persona bots — other bots in bots[] with a narrow role of their own (a secretary, a character). People talk to them directly; they are not the chief and not in its chain.
• Boss — Boss mode (boss-mode.md): a time-boxed run that answers HITL requests in the owner’s place. It stands in for the owner, not for the chief — it answers questions agents (the chief included) ask.
• System agents — agents the engine spawns for itself: git conflict resolvers, the OAuth authorizer.

Delegation is a convention, not a mechanism: the chief may call any agent directly.

People

Full list: chat-access.md § Who is who.

• Owner — a person with the owner tier; there can be several. Primary owner — the first id in TELEGRAM_ALLOWED_IDS, who gets the defaults.
• Member / stranger — the other tiers. Counterpart — the other side of a secretary chat.
• Admin — whoever runs the server. Brain author — whoever writes the brain, usually an agent.
• Requester — who asked for something (a person or an agent). Responder — who answered a HITL card.

What a person is shown

• Request (a HITL card) — a decision a person makes and something waits on: an approval, an answer, a grant (human-input.md).
• Artifact — a published file at a link, read in a browser (artifact-publishing.md). “Page” in these docs means the HTML a person opens — a publication’s, or a widget’s.
• Widget — a live interactive page an agent, pipeline or hook opens for a person, talking back to the brain while it is open (widget_*, widgets.md). Never a blocking decision — that is a request.