Workflows & Agents
How conversation flows are defined in Lemu
In Lemu, what you see as an agent in the dashboard is called a workflow in the API. They are the same thing — a workflow is the underlying definition, agent is the product name for it.
Anywhere the API says workflow, think "agent". Anywhere the API says workflow_definition, think "the conversation logic inside your agent".
The graph model#
A workflow is a directed graph — a set of nodes connected by edges.
- 1
Start Call
- 2
Qualify Intent
- 3
Then one of
Wants supportSupport Agent
Wants salesSales Agent
- 4
End Call
Nodes are the steps in the conversation. Each node has a prompt that tells the LLM what to say and do at that point.
Edges are the transitions between nodes. Each edge has a condition — a natural language description of when to move on. The LLM evaluates whether the condition has been met based on the conversation so far.
Node types#
| Type | What it does |
|---|---|
startCall | Entry point for telephony calls. The first thing the agent says when a call connects |
agentNode | An LLM-powered conversation step. The core building block |
globalNode | Defines instructions that apply across all agent nodes (e.g. tone, language, fallback behaviour) |
endCall | Terminates the call |
trigger | Entry point for API-triggered runs (non-telephony) |
webhook | Fires an HTTP request when reached — use for CRM updates, notifications, etc. |
qa | Runs automated quality analysis on the completed call |
Edges and transitions#
An edge connects two nodes and fires when its condition is satisfied:
transition_speech is optional — if set, the agent speaks it before moving to the next node.
Versioning#
Every time you update a workflow's workflow_definition, Lemu saves a new version while keeping the history. The current version is always what runs. Old versions are retained for auditing.
Creating workflows#
There are two ways to create a workflow via the API:
- From a definition — provide the full node/edge graph yourself. Best for programmatic generation.
- From a template — describe the use case in natural language and Lemu generates the initial graph using an LLM. Best for getting started quickly.