> ## Documentation Index > Fetch the complete documentation index at: https://docs.erstan.com/llms.txt > Use this file to discover all available pages before exploring further. # Core concepts > The mental model behind Erstan: workspaces, teams, agents, skills, tasks, connectors, knowledge, and approvals. Erstan puts AI agents to work on your NetSuite finance operations — drafting, analyzing, and acting on your data, with you in control of every change. This page explains the handful of building blocks that make that work, and how they fit together. Read it once and the rest of the docs will click into place. Each primitive links to its own deep-dive page. ## The building blocks A company account bound to one NetSuite environment (production or sandbox). You can belong to several workspaces and switch between them. Everything else lives inside a workspace. A focused container inside a workspace — Finance, Sales Ops, or the workspace-wide **Company** team — holding its own tasks, content, agents, and members. An AI worker that does a job for you. Built from a trigger, steps, instructions, skills, and connector tools. Three shapes: **Assistant**, **Workforce**, and **Automation**. A reusable package of instructions (a `SKILL.md` plus optional files and actions) an agent loads to do a task the same expert way every time. Published skills also show as **Agent ready**. A unit of agent work you can track. Every run becomes a task that moves through **Queued → In Progress → Needs Review → Complete**. A connection to an external system — NetSuite (flagship), LLM providers, Slack, MCP servers, and built-in Erstan tools — that gives agents the access and tools they need. Retrieval sources agents can pull from: searchable **Team Files** (**Agent knowledge**) and Erstan-managed **NetSuite product updates**. A human-in-the-loop pause. Before an agent writes to NetSuite or another system, it stops for you to **Run action** or **Don't run**. ## The three agent shapes An **Agent** is the AI worker you chat with, build, and run. Every agent comes in one of three shapes, depending on the job. | Shape | Also called | What it does | How it starts | | -------------- | ------------- | --------------------------------------------------------------------------------------------- | ---------------- | | **Assistant** | Domain expert | A chat-first agent that plans, retrieves NetSuite data, and acts, looping you in when needed. | Chat | | **Workforce** | Orchestrator | Routes work across assistants, tools, and human handoffs, then recombines the results. | Chat | | **Automation** | Set process | A multi-step process that runs the same way every time, with approval gates. | Form or Schedule | The product calls all three **Agents**. Under the hood, Assistants and Workforces are chat-first agents and Automations run as a fixed sequence — but you rarely need to think about that distinction. See [Agents overview](/agents/overview) for the full picture. ## How they fit together The pieces stack up in a simple chain. A connector grants access, an agent does the work using skills and tools, the work shows up as tasks, and you approve anything that writes back to NetSuite. Your workspace owner [connects NetSuite](/connectors/netsuite) once, and each member [authorizes their own NetSuite user](/connectors/netsuite-authorize). A [connected LLM provider](/connectors/llm-providers) supplies the model that powers chat and agents. You [chat with an agent](/chat/overview) or [run one](/agents/run-and-tasks). It uses its [skills](/skills/overview) for know-how and connector [tools](/agents/tools-skills-connections) to read NetSuite data — for example, finding overdue invoices or summarizing AR aging. It can also pull from [Knowledge](/content/agent-knowledge) when it needs your files or NetSuite product updates. Each run becomes a [task](/tasks/overview) you can follow from **Queued** through **In Progress**, **Needs Review**, and **Complete** — viewable as a list or a Kanban board. Before an agent writes to NetSuite — creating a record, sending a payment, editing a transaction — it pauses for an [approval](/chat/approvals). You review the change and decide whether it runs. ## NetSuite, safely Because NetSuite is the system of record, Erstan treats every write with care. Agents act in NetSuite **as you**, using your own NetSuite role and permissions ([per-user authorization](/connectors/netsuite-authorize)). Each tool has a **write policy** — **Allow**, **Require approval**, or **Deny** — and write-capable actions require approval by default. **Auto-approve writes** can skip the prompt; turn it on only when you fully trust the agent and the action. See [Security & write safety](/security/overview). A workspace binds to one NetSuite environment. Production and sandbox each get their own workspace, so test work never touches live data. When NetSuite is under load, a brief **"NetSuite is busy" queue** can hold work and resumes automatically — no action needed. ## Where each piece lives Switch workspaces from the workspace switcher. Inside a workspace, **Teams** organize people, agents, content, and tasks. The **Company** team is workspace-wide; you can also create focused teams like Finance or AP. Roles are **Owner / Admin / Member**. Build and run agents from the [Agents library](/agents/overview). Reusable instruction packages live under [Skills](/skills/overview). Ready-made agents, skills, and bundles come from the [Marketplace](/marketplace/overview). Track agent work in [Tasks](/tasks/overview); review outputs in the **Needs Review** and **Approvals** queues. Write actions surface as [approval cards](/chat/approvals) in chat and on runs. Manage external systems under [Connectors](/connectors/overview). Files and documents live in a team's **Content**; make a file searchable to turn it into [Agent knowledge](/content/agent-knowledge). ## Next steps Connect NetSuite and run your first agent. Find your way around the sidebar and Home. Go deeper on the three agent shapes.