Vocabulary

Company & team

Company

A tenant. The unit that owns AI Employees, Connections, Channels, Notes, Bases, and so on. One Genosyn install can host many.

Member

A human user inside a company. Roles: owner, admin, member. Don't say 'User' in product copy — that's reserved for the DB entity name.

Team

A subgroup of members for routing and notifications.

AI substrate

AI Employee

A persistent AI teammate attached to a company. See AI Employees.

Soul

The employee's written constitution. Markdown on AIEmployee.soulBody. See Soul.

Skill

A reusable playbook. Markdown on Skill.body. See Skills.

Routine

A scheduled, recurring piece of work. Cron-triggered. See Routines.

Run

One execution of a Routine. The agent transcript is captured on Run.logContent (256 KB cap).

Trigger

An event subscription that fires a Routine when a resource family changes — id-only frames, so an event routes work but never carries content. Never "Subscription," "Listener," or "Hook" — and a Revenue Signal stays a cron-evaluated query, never a Trigger. See Reactivity.

Wakeup

A timed follow-up session an AI Employee schedules for itself, with a note its future self reads — at most 20 pending, at most 90 days out. Never "Reminder," "Timer," or "Snooze." See Reactivity.

Workstream

A persistent state document for work spanning many Runs, maintained by the employee and opened into every Run brief of the Routine it binds. Never "Thread" or "Epic" — and "Project" stays reserved for the humans' task manager. See Reactivity.

Initiative

Standing work an AI Employee proposes — evidence, case, and the exact Routine it wants — that exists only once a human accepts it. Never "Suggestion" or "Idea" — and "Proposal" belongs to Revision proposals. See Reactivity.

AI Model

A brain an AI Employee runs on — an Anthropic or OpenAI API connection, a custom OpenAI-compatible endpoint, or a trusted single-tenant OpenAI subscription connection through the official Codex app-server. An employee can hold several and keep one active. See AI Models.

Handoff

A formal AI→AI delegation with a status workflow — open, accepted, completed, cancelled.

Integration words

Integration

A connector type. Static catalog defined in code under server/integrations/providers/.

Connection

One authenticated account inside an Integration. DB row, per-company.

Grant

An AI Employee's access to a specific Connection.

MCP server

A Model Context Protocol server. Genosyn ships two built-ins (genosyn, browser) and any number of user-registered ones via the McpServer entity.

Workspace surfaces

Channel / DM

Slack-style workspace chat between humans and AI. WebSocket-backed, mentions auto-invite the employee.

Notebook / Note

Notion-style company-wide markdown knowledge base. Tree-structured, soft-deletable.

Base

Airtable-style multi-table workspace with views, comments, attachments.

Pipeline

DAG of typed nodes for deterministic glue. Distinct from a Routine — Routines are AI-driven, Pipelines are wire-driven.

Project / Todo

Tasks (the human kind). Projects hold Todos; Todos have comments. Distinct from Routines.

Resource

External material an employee studies — articles, ebooks, transcripts. Distinct from Notes (team-authored) and Memory (atomic facts).

Chart / Dashboard

Explore — Metabase-style BI over database integrations. Save SQL as a Chart, pin Charts onto a Dashboard.

Control surfaces

Approval

A gate that blocks an action until a human ✓. Kinds include routine, browser_action, mcp_tool, and ad_spend.

Budget

A monthly envelope over authorized ad-spend increases, scoped to the company, one Connection, or one AI Employee — every applicable envelope must have headroom, and the tightest binds. Enforced on every spend-increasing path, approval replays included; never blocks a spend decrease. See Company policies.

Check

A machine-verifiable assertion a Run must pass before it finalizes green — a predicate over the effects the server recorded, or a command that must exit 0 in the sandbox. Written by an operator, evaluated by the server, and unreachable from every MCP tool. Never a "test," an "assertion," or a "gate" — System Health keeps "probe" for its own diagnostics. See What proves a Run worked.

Decision

A question an AI Employee stopped to ask, with the options it will act on. The employee raises it and a Member answers; unlike an Approval, nothing is executed on your behalf. See Decision stack.

Goal

A measurable objective — target value, direction, optional deadline, optional owning AI Employee — cascading company → employee. Humans set Goals; employees read them in every prompt and report progress. Never "OKR," "KPI," or "Target." See Goals.

Lesson

What a failed, timed-out, off goal, or Check-failing Run teaches the next one: a cause and an advice, written by a restricted retrospective turn, opening the Routine's future Run briefs until dismissed. Never "Learning," "Insight," or "Retro." See The improvement loop.

Policy

A company-wide rule binding every AI Employee at once: prose injected above every Soul, blocked recipient domains refused at the mail-send choke point, and forbidden tools refused at dispatch — each refusal a policy.violation audit event. See Company policies.

Revision proposal

A complete replacement body an AI Employee stages for its own Soul, a Skill, or a Routine's brief or acceptance criteria, with a rationale and evidence Runs. Nothing changes until an owner/admin applies it from the Revisions page. See The improvement loop.

Standdown

A revocable stop on all AI work at company, employee, or Routine scope, placed by an admin or by the consecutive-failure breaker. The exact inverse of a Waiver — imposed rather than earned, broad rather than narrow — and distinct from Routine.enabled, which stays the ordinary switch. Never "pause," "hold," "suspend," or "freeze." See Standdowns.

Waiver

One approval gate switched off for an AI Employee that earned it — browser submits for the employee, or gated ticks for one Routine. Proposed by the eligibility sweep through the Approvals inbox; revoked automatically by any Run that failed, timed out, graded off goal, or failed a required Check. Never "trust score," "tier," or "level." See Earned autonomy.

Audit event

Append-only log of every consequential action. Used for after-the-fact review and the activity feed.

Journal

Append-only diary an employee writes about their own work. AI-only writes; humans read.

Notification

An item in the inbox — mention, approval needed, run failed, etc.