Getting started
Core concepts & glossary
Every term you meet in Turtle AI Coworker, what it means, and how the pieces fit together.
Turtle AI Coworker has a small vocabulary that comes back on every screen. This page explains how the pieces fit together, then defines every term in one place, grouped by area.
Use it as a reference. When a module page uses a word you have not met, look it up here.
How the pieces fit together
Work starts from a person or a trigger. A worker picks it up, reasons with a model, and uses the tables, knowledge bases and tools it was given. Governance checks each tool call before it runs, and the whole run is recorded.
StartRun page, chat, trigger
Workeragent, team, employee
Resourcestables, knowledge, tools
Governancepolicy, approval, budget
Run logthe audited record
Where things live
An organization contains workspaces. Workers, tables, knowledge bases and dashboards live in a workspace. Tools and models are connected at the organization level, then granted to workers. Members belong to the organization with a role, and can be put in groups for sharing.
| Term | Type | What it does |
|---|---|---|
Organization | Tenancy | The top level: your company. It holds members, workspaces, connected tools and models, and billing. You can belong to more than one organization and switch between them. |
Role | Tenancy | What a member may do in an organization: Owner, Admin, Builder or Member. Tools, Models, Users, Groups and Compliance are shown to owners and admins only. |
Workspace | Tenancy | A separate area inside an organization with its own agents, teams, employees, tables and knowledge bases, for example one per department or one per client. Most screens work on the workspace you have selected. |
Group | Tenancy | A named set of members. You share agents, teams, employees, tables and knowledge bases with a group instead of one person at a time. |
Workers
There are three worker shapes. A sequential agent does one job. A team answers a conversation with several specialists. An AI employee owns a role and works through a queue. Teams and employees reuse your agents, so you can start small and grow.
| Term | Type | What it does |
|---|---|---|
Sequential Agent | Worker | One worker that does one job end to end: typed inputs, an ordered list of tasks, scoped access to tables, knowledge bases and tools, and a recorded result. |
Task | Agent part | One step of a sequential agent, with a title, a description (the instruction) and an expected output. Tasks run top to bottom and a task can use the inputs and earlier outputs. |
Input parameter | Agent part | A typed value the caller supplies when the agent runs, such as short text, long text, number, URL, select, boolean, date or file. You reference it in tasks with {{ key }}. |
Agentic Team | Worker | One chat backed by several member agents. You talk to the team as a whole, and each message goes to the member best suited to it. |
Supervisor | Team part | The part of a team that reads each message and routes it to a member agent, or answers directly, following the team's routing plan. |
AI Employee | Worker | A named coworker that owns a role. It follows standing instructions, keeps memory, plans multi-step work and uses the agents, tables, knowledge bases and tools you give it access to. |
Task queue | Employee part | The list of work an AI employee has been given, with a status: draft, pending, running, waiting for a person, completed, failed or cancelled. Tasks come from chat, schedules or events. |
Memory | Employee part | What an AI employee keeps between conversations: facts, preferences, standing instructions, learnings, relationships and summaries of finished work. It is used to write better answers next time. |
Data and knowledge
Tables are for records a worker reads and writes. Knowledge bases are for reference material a worker searches. Interfaces put tables in front of people.
| Term | Type | What it does |
|---|---|---|
Table | Data | Structured records with typed columns that people and workers read and write. Workers get independent Read, Create, Update and Delete permissions per table. |
Column type | Data | The kind of value a column holds: Text, Number, Date, Select, Boolean, URL, Email, File, User, Foreign Key (a link to a row in another table), Agent Output and AI Enriched. |
AI-enriched column | Data | A column that the platform fills in by running a worker on the row, instead of a person typing the value. Its value cannot be set by an outside caller. |
View | Data | A saved way of looking at a table: Grid, Kanban, Calendar, Gallery or Grouped List, with its own filters and sort. |
Knowledge Base | Data | A searchable collection of reference material. Workers search it by meaning while they run and never write to it. |
Source | Data | One item added to a knowledge base: an uploaded file, a website, pasted text, or content synced from an app such as Google Drive, Notion, Confluence or Slack. |
Interface | Data | A dashboard built on your tables: pages of blocks such as metrics, charts, record lists, forms, approvals and chat. You can share it with the organization, with chosen people or through a public link. |
Connections and activation
| Term | Type | What it does |
|---|---|---|
Tool | Action | A set of actions a worker can take in another app, such as sending an email or updating a CRM record. Each action a worker takes is a tool call. |
Integration | Action | A connected account for a tool. An owner or admin connects it once for the organization, and it is then granted to workers one by one. |
Custom tool | Action | Your own REST endpoints, grouped into a collection, that workers can call like any other tool. |
Custom MCP server | Action | Your own Model Context Protocol server, added by URL with no auth, a token, a custom header or basic auth. Its tools are read when a worker runs, so new tools on your server need no change here. |
Model (BYOK) | Reasoning | The language model a worker reasons with. You bring your own key for a provider and pick the model per worker. BYOK means bring your own key. |
Azure deployment | Reasoning | A model you have deployed in your own Azure OpenAI resource. You connect the resource and add each deployment with the base model it runs. |
Trigger | Activation | What starts work without a person pressing Run: a schedule, an event in a connected app, a webhook, or a table row being created or changed. |
Packs, templates and the Assistant
| Term | Type | What it does |
|---|---|---|
Solution Pack | Bundle | A pre-wired department installed in one step: tables, knowledge bases, agents, teams, an AI employee, triggers, and often dashboards and policy rules. |
Template | Bundle | A single agent, team or employee you import from the library into a workspace, with the resources it needs. |
The Assistant | Setup | The in-app chat that sets up your workspace: it researches your company, proposes packs, helps you connect models and tools, installs, and proposes triggers and setup plans. It builds, it never runs workers. |
Runs, governance and testing
Governance answers three questions about every worker:
- What can it do? Per-table permissions, tool grants, policy rules and budgets.
- What did it do? The run log and audit trail of every tool call, with secrets redacted.
- Who approved it? Approvals, enforced at the moment the tool call is attempted.
| Term | Type | What it does |
|---|---|---|
Run | Execution | One execution of a worker, with its inputs, steps, tool calls, output, duration and cost. |
Run log | Execution | The stored record of one run: who or what started it, the status, the totals and the list of tool calls made during it. |
Audit trail | Control | The history of every run and every tool call, with inputs and outputs stored after secrets are redacted, plus the policy decisions and approvals that applied. |
Policy rule | Control | A workspace rule matched against tool calls before they run. Its action is Allow, Deny, Require Approval, Alert or Log Only. |
Approval | Control | A request for a person to approve or deny a held tool call. An approval covers that one call; a reviewer can correct values before approving. |
Circuit breaker | Control | An automatic brake on repeated failures, with three levels: yellow alerts, orange pauses and waits for review, red halts. |
Budget | Control | A daily or monthly spend cap for a workspace. A soft cap sends an alert; a hard cap refuses new runs once reached. |
Eval suite | Testing | A named set of test cases for one sequential agent or agentic team. |
Eval case | Testing | One input plus the outcome a correct run must achieve, written as a condition a person could check. |
Eval run | Testing | One pass through a suite. Each case runs for real, is graded, and is compared with its previous result. Eval runs are not counted against your plan's runs. |