Glossary

A quick reference for the vocabulary of Syntrum. Terms are grouped by theme; each links to the page where it's explained in full.

Core model#

Agent: An AI employee you hire. It has an identity (name and role), an objective, responsibilities, a communication style, and a cadence, and it runs continuously — watching for work and acting on it. See What is Syntrum.
Digital workforce: Syntrum's core idea: instead of building automations, you hire AI employees and manage them like a team.
Workspace: Your company's shared space in Syntrum — one workspace = one company. Agents, integrations, members, policies, and billing all belong to a workspace. See Agents, workspaces & integrations.
The agent loop: The five-phase cycle every agent repeats: Observe → Understand → Decide → Act → Report. See The agent loop.
Integration / tool / toolkit: A connected external app (Gmail, Slack, etc.) — an agent's senses and hands. A toolkit is the bundle of actions a given tool exposes. Connected per user via OAuth. See Connecting tools.
Attach / detach: Granting a toolkit to an agent (attach) or revoking it (detach). Attaching = capability granted. An Admin-and-above action.
Memory: What an agent remembers — past decisions, learned tool skills, and guidance turned durable — so it improves over time. See Memory & learning.
Heartbeat: The steady tick of a running agent's loop — the rhythm at which it checks for and acts on new work.
Boot / starting-up: The brief state while an agent's runtime environment ("pod") starts before it can act. Shown as a calm boot animation, not an error. See Troubleshooting.

Governance — the three planes#

The three planes: The mental model for every permission question: human RBAC, agent autonomy, and tool permissions. See The three planes.
Cascade (Workspace → Agent → Tool): How every policy resolves — the most specific setting wins, and levels inherit by default. See The three planes.
Inherit: The default at every level — a setting follows the level above it until you deliberately pin it. "Sync unless pinned."
Pin: A deliberate override at the agent or tool level that stops it inheriting. One click from Reset to workspace default.
Block is a floor: A core rule: a block set at a higher level can be tightened but never loosened downstream. Approving a related action can't override a block. See Approvals.
Supervision mode: An agent's autonomy setting: Inherit, Supervised (every side-effecting action needs approval), Balanced (workspace categories apply as-is), or Autonomous (approval only where the workspace requires it). See Supervision & modes.
Agent autonomy: How much an agent may do without a human — set by its supervision mode plus the sensitive-action categories. See Agent autonomy.
Guardrail: A workspace-level rule over sensitive action categories (payments, destructive, data export, user management, external messaging, security links), each set to allow, require approval, or block. This is the floor every agent inherits. See Guardrails & sensitive actions.
Tool dial / tool permission: The per-toolkit policy on an agent — Off / Read-only / Asks first / Autonomous (or Inherit) — that says what an agent may do with one tool. See Tool permissions.
Access level: Part of a tool permission: whether an agent can read, write, or take destructive actions with a tool.
Oversight pin: A pin that raises the human-oversight requirement for a category or tool (e.g. forcing approval), independent of the base access level.
Provenance: The record of which rule, at which level (workspace / agent / tool) decided an action — shown on approvals and in the audit log as "Held by workspace policy / a tool pin / agent policy." See Approvals.

Operating#

Inbox: Where you and your agents converse — direct chats and escalations. See Chatting with agents.
Conversation / thread: A single topic of discussion with one agent. A chat is one you start; an escalation is one the agent starts.
Escalation: A conversation an agent opens because it wants your input on a situation — carrying risk level, impact, cost, and affected systems. You answer with guidance. See Chatting with agents.
Guidance: Instruction you give an agent in the Inbox to steer how it handles a situation — as opposed to approving one specific action.
Quote-reply: Replying to a specific agent message so the agent knows exactly which message you answered.
Approval: A single agent action, paused, waiting for a human yes/no because a rule requires it. Resolve it in the Approvals queue — Operator and above.
Edit before approve: Correcting an approval's action payload before letting it run. Your edited version executes; the original is kept in the audit trail. See Approvals.
Action payload: The literal, complete description of what an agent proposes to do in an approval — recipients, amounts, content.
Risk level: A rating on approvals and audit events — Low / Medium / High / Critical — to help you triage.
Take over / human-in-control: Temporarily driving an agent yourself; gated actions then execute as your decision and are audited as such.

Insights#

Report / Performance Report: The live-data review of your workforce — output, success rate, time saved, cost, and safety, over 7/30/90 days. See Reports.
Success rate: Share of finished agent work that succeeded without intervention.
Time saved: Estimated manual work your agents avoided, versus doing it by hand.
Audit log: The immutable record of every action, decision, learning event, and guardrail hold, with provenance. See Audit log.
Insight: An automatically surfaced pattern across your agents' work, with a recommendation and a link to act on it.

Administration & security#

Role (RBAC): A member's permission level — Owner / Admin / Operator / Viewer. See Team & roles.
Owner: Full control including billing, ownership transfer, and workspace deletion. One per workspace.
Admin: Manages members (below their own role), agents, tools, and policy — but not billing or ownership.
Operator: Runs the workforce day to day — chat, approvals, start/stop — but can't change the rules.
Viewer: Read-only access.
Transfer ownership: Handing the Owner role to another member in one atomic step (you become an Admin). Owner-only. See Team & roles.
Presence: Whether a member is online (active in the last five minutes) or when they were last active.
BYOK (Bring Your Own Key): Supplying your own AI provider API keys so agents bill to your accounts. Keys are stored encrypted; only the last four digits are shown. See Models & BYOK.
Default model: The AI model every agent action runs on. See Models & BYOK.
Backoff model: The automatic fallback model used when the default is rate-limited or down. Can't equal the default.
Two-factor authentication (2FA): A second sign-in step — a rotating 6-digit code from an authenticator app. See Authentication & 2FA.
Recovery code: A one-time code that signs you in if you lose your authenticator. Shown once when you enable 2FA.
Session: A record of a signed-in device. Review and revoke them, or sign out everywhere. Grouped by device. See Sessions & devices.
Message toast: An in-app pop-up (bottom-right) announcing a new agent message; click it to open the thread. See Notifications.
Desktop notification: A browser-level alert (requires granting browser permission) for events even when the tab isn't focused.

What's next#

Troubleshooting — fixes that use this vocabulary.
The three planes — the governance model tying many of these terms together.
Introduction — the big picture, if you're just arriving.