Core concepts & glossary
The vocabulary used throughout these docs, and how the pieces connect into one governed system.
Turtle AI Coworker has a small vocabulary that recurs everywhere. This page defines each term and shows how they fit together, so the rest of the documentation reads cleanly.
How the pieces fit together
Work enters through a trigger, a worker takes it, the worker reasons and acts on your data and tools using a model, governance bounds what it may do, and the whole run is traced. That single loop is the product.
The three worker shapes
Pick the worker that matches the job. You can start with one agent and grow to a whole employee without rebuilding, because each shape is built from the one below it.
- Sequential Agent for one repeatable job with a clear start and end.
- Agentic Team when one conversational surface must handle several different intents, each needing a different specialist.
- AI Employee when you want to delegate a whole role that persists, remembers, and runs work in the background.
The shared substrate
Whichever worker you build, it draws on the same four building blocks, and every one of them is configurable and permissioned.
- Knowledge Bases ground answers in your documents through retrieval, and cite what they used.
- Tables are the structured memory workers read and write, so results become durable data rather than vanishing chat.
- Tools & Integrations let workers take real actions in the apps you already run.
- Models are the reasoning engine, chosen per agent, always on your own keys.
Governance, in one line each
Governance answers three questions about every worker, by design rather than as an afterthought:
- What can it do? Per-table read, create, update and delete; per-run caps; tool scopes.
- What did it do? A full audit trail of runs and tool calls, with sensitive data flagged and secrets redacted.
- Who approved it? Human-in-the-loop approval gates, enforced at the moment a write is attempted.
Glossary
Every term you will meet in these docs, in one place.
| Term | Type | What it does |
|---|---|---|
Sequential Agent | Worker | A single worker that runs one job end to end: typed inputs, ordered tasks, scoped access, a traced result. |
Agentic Team | Worker | A chat interface backed by several sub-agents. Each message is routed by intent to the right specialist. |
AI Employee | Worker | A named coworker you hire and delegate to. Owns a role, keeps memory, plans multi-step work, asks approval on risk. |
Table | Data | Structured business memory. Typed columns, including AI-enriched columns that fill from agent runs. Workers read and write rows under least-privilege permissions. |
Knowledge Base | Data | A vectorized set of documents a worker searches by meaning at run time, then cites. Read-only to workers. |
Tool / Integration | Action | A scoped action in one of your apps (send an email, update a record). Connected once, permissioned per worker. |
Custom Tool / MCP | Action | Any REST endpoint or MCP server turned into a tool the workforce can call. |
Model | Reasoning | The LLM behind a run. Provider-agnostic, chosen per agent, with your own API keys. |
Trigger | Activation | What starts a run: a schedule, an integration event, a webhook, a table-row change, an upstream agent, or a threshold. |
Run | Execution | One execution of a worker. Every run has inputs, steps, tool calls, writes and a result. |
Trace | Observability | The recorded log of a run: what it read, which tools it called, what it wrote, and who approved it. |
Governance | Control | The layer over every run: permissions, approval gates, budgets and the audit trail. |
Solution Pack | Bundle | A complete working department (tables, agents, an AI employee and triggers) installed in one step. |
Workspace | Tenancy | An isolated environment with its own data, workers, members and permissions. Organizations can have more than one. |