Core Concepts

This section explains the building blocks you compose when you build on Ema. Read it before you open the AI Employee builder — once these concepts click, every screen in the product maps cleanly onto one of them.

An AI Employee (AIE) is the top-level unit you build, configure, and deploy. Everything else on this page is something an AI Employee is made of:

• A workflow — a directed acyclic graph (DAG) of typed nodes — defines what the AI Employee does.
• Agents are the reasoning nodes inside that workflow. You configure each from one of Ema's agent types, built from templates.
• Knowledge bases give agents private, searched context through retrieval-augmented generation (RAG).
• Human in the loop (HITL) pauses a run for a person to approve, fill in a form, or answer a question, then resumes.
• Conditions and expressions route a run down different branches and pull data between nodes using {{variable}} references.
• Versioning separates the draft you edit from the published version end users run, and keeps an immutable history you can promote or revert to.
• An Orchestrator AI Employee has no workflow. It coordinates other AI Employees from a conversation, using Instructions and a list of connected AI Employees you configure.

How the pieces fit together

When an AI Employee runs, the platform:

1. Receives an input (a chat message, a dashboard row, an API call, a document, or a schedule fire) and maps it to the workflow's input_params.
2. Walks the workflow DAG in topological order, executing each node when its inputs are ready. Nodes with no dependency on each other run in parallel.
3. For each agent node, calls the agent runtime, which performs LLM reasoning, optional knowledge base retrieval, and optional Tool calls.
4. Evaluates conditions on the edges leaving each node to decide which downstream branches run and which are skipped.
5. Pauses on any human in the loop node until a person responds, then resumes from where it left off.
6. Collects the outputs declared by the workflow's publish nodes and returns them, recording a full step-by-step trace.

The reasoning itself is powered by EmaFusion™, Ema's model-routing layer, which picks the best large language model for each subtask. See EmaFusion™.

What's in this section

• AI Employees — the top-level container: its lifecycle, the resources it owns, and the surfaces it deploys to.
• Orchestrator AI Employees — the interaction type that delegates to other AI Employees instead of running a workflow: what you configure and how it behaves in chat.
• Workflows — the DAG model: node types, edges, parallelism, drafts and versions, runs and steps.
• Agents — the agent types and how each one is configured.
• Knowledge Bases — folders, chunking, embedding, and how agents retrieve context with RAG.
• Human in the Loop — pausing and resuming a run for approval, a form, or a question.
• Conditions and Expressions — {{variable}} references and the full operator catalog for branching.
• Versioning — drafts versus published versions, publish immutability, version history, and what happens to in-flight runs when you publish.

Where to go next

Once these concepts are familiar, move on to the hands-on material:

• Builder guides — step-by-step instructions for building in the AI Employee builder.
• Agent reference — per-agent configuration detail.
• Autopilot — build and edit AI Employees by describing what you want in plain language.