What is SuperFlow

SuperFlow is Lyzr’s visual workflow builder. It lets you chain AI agents, LLM calls, tools, HTTP requests, code, conditionals, loops, and human approvals into a single automation on a drag-and-drop canvas, with no glue code to maintain. SuperFlow is built for production workloads. Every step is durably journaled, every successful step runs exactly once, and every in-flight run survives crashes, restarts, and deployments without losing work. Long-running approvals can pause for days or weeks at zero cost. Failed steps retry automatically. Past runs are replayable end-to-end. You can find SuperFlow in Agent Studio.

When to use SuperFlow

A single Lyzr agent handles tasks that fit one model with a set of tools. SuperFlow is the right choice when:

The workflow involves multiple steps that are not all LLM calls, for example: fetch from an API, run an agent, branch on the result, call another agent.
The flow requires branching or looping based on data (if/else, case switching, iterate over a list).
A human needs to approve something mid-flow before the workflow continues.
You want to schedule the workflow to run on a cron (every Monday morning, every 5 minutes, and similar schedules).
You are composing several agents together, such as a researcher feeding a writer feeding an editor.

Mental model

A SuperFlow is a graph of nodes connected by edges:

Every node has inputs and outputs, and data flows along the edges as JSON.
Each node receives the output of the nodes upstream of it, does its job, and emits its own output downstream.
A SuperFlow always starts with one Trigger node, which defines what input the workflow accepts.
Downstream nodes can reference any upstream node’s output using expressions like {{ $('Trigger').json.email }}.

This is the same idea as a flowchart, except every box is an executable step and the lines between them carry real data.

The SuperFlow editor

A new SuperFlow opens with a Trigger node already on the canvas. The SuperFlow editor has three areas. The node library on the left lists every node you can add, the canvas in the middle holds the flow, and the Build with AI panel on the right lets you create or change the flow by describing it.

Node library

The node library groups nodes by category, such as AI, Safety, Control Flow, Data, and Apps. Select a category tab to browse its nodes, or type in Search nodes to find a node by name. To add a node, drag it onto the canvas or double-click it. See Node types for the nodes in each category, and the node reference for how each node works.

Build with AI

Build with AI creates a flow from a plain-language description. It designs the flow from SuperFlow’s nodes, draws each node on the canvas as it goes, and adds a plain-language summary of what the flow does. You can start a flow this way and then refine it by hand, or keep working in the chat. To build a flow with AI:

In the Build with AI panel, describe what the flow should do in a sentence or a paragraph. For example, “Read prospects from a Google Sheet, research each one, draft an outreach email, and hold it for approval before sending.”
To point the flow at resources you already have, select + and attach an agent, knowledge base, guardrail policy, SuperFlow, or tool for it to use.
Review the flow on the canvas and read the summary.
Select Run sample input to test the flow end to end before you deploy it.
To change the flow, describe the change in the same chat, for example “Add an approval step before the email is sent, and route rejections back to the drafting node.”

Each chat message is undone as a single step. Press Ctrl+Z to undo the last change Build with AI made.

Create an agent from the editor

When a flow needs an agent that does not exist yet, you can build it from the editor and use it in the flow straight away.

Add an AI Agent node to the canvas and open it.
Under Use Agent, select Create agent. The agent builder opens, with Back to SuperFlow in the top right if you want to return without creating the agent.
Set up the agent’s role, goal, instructions, model, and any knowledge, tools, skills, or features it needs. The builder works the same way as the Studio Builder.
Select Create & use. SuperFlow creates the agent and uses it in the AI Agent node.

To change an agent that a node already uses, select Edit agent on the node. Changes to an agent apply everywhere that agent is used.

Built for production workloads

Durable execution. Every meaningful step (LLM call, tool call, HTTP request, code execution, loop iteration) is journaled to durable storage before completion. A service restart in the middle of a 50-step workflow resumes from the exact point of failure, not from the beginning. Exactly-once side effects. A step that has already run successfully will never run again. No double charges, no duplicate emails, no re-fired API calls on retry. Once the journal records a step as completed, the engine treats it as final.

Durable execution. Every meaningful step (LLM call, tool call, HTTP request, code execution, loop iteration) is journaled to durable storage before completion. A service restart in the middle of a 50-step workflow resumes from the exact point of failure, not from the beginning.
Exactly-once side effects. A step that has already run successfully will never run again. No double charges, no duplicate emails, no re-fired API calls on retry. Once the journal says “this step completed,” the engine treats that as final.
Pauses are free. A SuperFlow waiting on a human approval, a schedule, or a delay consumes essentially no resources. Runs can wait hours, days, or weeks and resume exactly where they paused.
Crash-safe schedules. Cron triggers use durable timers, not in-memory schedulers. A schedule won’t “miss a tick” because the runtime was down.
Per-node retries with exponential backoff. Configure every node that talks to an external system to retry on transient failures. Retries are themselves durable: the retry count survives restarts.
Error edges for graceful failure. When retries are exhausted, opt into an error output that routes failures to a recovery path (a fallback API, a Wait for Approval for human escalation, or a notification) instead of failing the whole run. Strictly opt-in per node.
Live and historical observability. Watch a run in real time on the canvas; replay every past run with full per-node outputs preserved. OpenTelemetry traces are emitted end-to-end.

Triggers

Every SuperFlow starts with a Trigger node. Four modes are available:

Mode	Use
Manual	Click Run in the editor. Best for testing.
Webhook	An external system posts to a URL and the payload becomes the Trigger input.
File Upload	Someone uploads a file and the run receives it as `file_url`, `file_name`, `file_type`, and `file_size`. Pair with a Parse node to extract the file’s contents.
Schedule	Runs automatically on a cron expression using crash-safe durable timers.

Webhook example:

curl -X POST "https://<your-webhook-url>" \
  -H "X-Webhook-Secret: <your-secret>" \
  -d '{"customer_id": "cus_123", "event": "trial_expiring"}'

Expressions

Downstream nodes reference upstream outputs using expressions. All expressions use the {{ }} syntax:

Expression	What it returns
`{{ $json }}`	The entire input item
`{{ $json.field }}`	A single field from the input
`{{ $('Trigger').json.email }}`	A specific field from a named node
`{{ $('AI Agent').json.output }}`	The output from an AI Agent node

Full JavaScript is valid inside {{ }}, including ternary operators, array methods, and string operations. Use the field reference picker (plug icon) to build expressions without typing paths manually.

Node types

Nodes are organized in a palette by category:

AI

Node	What it does
AI Agent	Runs a Lyzr agent with all its tools, KB, and memory. The full agent config carries over.
A2A Agent	Invokes an external agent via the Agent-to-Agent protocol.
LLM	Calls a model inline with a custom system prompt. Can orchestrate sub-agents via ReAct.
AI Swarm	Runs multiple agents in parallel and merges results.
Knowledge Base	Retrieves chunks from a Knowledge Base, or trains a Knowledge Base on new text, at any step in the flow.
Decide	Asks a model for a choice, a score, or a probability, using the Jev decision model.
Rank	Reorders a list of items by relevance to a goal, using the Jev decision model.

Apps

Node	What it does
App tools	Run one action in an app, such as Gmail, Slack, Google Docs, or SharePoint. Each app is its own node, and Custom tool calls one of your own tools. See App tools.

Control Flow

Node	What it does
If / Else	Binary branch. Rule mode (visual conditions) or AI mode (natural language).
Switch	Multi-way branch on a field value.
Loop	Iterates over a list, once per item or in batches.
Loop Until	Re-runs a region of the flow until a condition holds or a budget runs out.
Filter	Removes items that do not match a condition.
Merge	Combines outputs from parallel branches.
Stop and Error	Terminates the run with an error.
Wait	Pauses for a fixed duration.
NoOp	Passthrough node, useful for labeling branch points.

Safety

Node	What it does
Guardrails	Applies a Responsible AI policy to the data flowing through this step.

I/O and Compute

Node	What it does
HTTP Request	Call any URL (GET, POST, PUT, PATCH, DELETE). Built-in authentication (Bearer, Basic, API Key, OAuth2 client credentials, AWS SigV4, JWT, mTLS) and native support for sending or receiving files as items.
Code	Execute JavaScript inline (Monaco editor). Globals: `$input.all()`, `$json`, `$items`.
Execute Workflow	Run another SuperFlow as a sub-workflow.

Data Transform

Node	What it does
Set	Reshapes data by adding, renaming, or removing fields.
Aggregate	Combines multiple items into one.
Sort	Orders items by a field.
Limit	Keeps only the first N items.
Remove Duplicates	Deduplicates items on a key field.
Rename Keys	Renames field names.

Human-in-the-Loop

Node	What it does
Wait for Approval	Pause the run for a human decision, optionally with a typed form for the approver (text, number, boolean, dropdown, date, file). Submitted form values flow downstream merged into the item. Two output paths: approved (0) and rejected (1). Pauses are durable: no cost while waiting.

Document

Node	What it does
Parse	Extracts text from PDF, DOCX, and CSV files.
Extraction	Pulls structured data from unstructured text.
Label	Classifies or tags document content.

Utility

Node	What it does
DateTime	Parses, formats, or calculates dates.
Crypto	Hashes or encodes values.
XML	Parses or generates XML.

Running and monitoring

During execution, the canvas shows real-time node status:

Amber pulse: currently executing
Green check: completed successfully
Red cross: errored
Blue pause: awaiting human approval

The execution panel shows expandable cards for each node with full JSON output. Lifecycle controls for long-running workflows:

Pause: cooperative pause at node boundaries
Resume: continues from the stopping point
Terminate: cancels the run permanently

All controls are durable and survive service restarts.

Sub-workflows and sub-agents

Sub-workflow: use the Execute Workflow node to call another SuperFlow as a step. The sub-workflow receives the current node’s output as its Trigger input. Sub-agent: toggle “Run as sub-agent” on an LLM or AI Agent node. The parent LLM can call the sub-agent as a tool during its ReAct loop. Sub-agents are connected with purple edges.

Overview

Agent Building

Automation

Orchestrate

Knowledge

Connections

Voice Agents

Safety and Evaluations

Monitoring

Governance

Versioning & Git

Blueprints

Lyzr App Store

What is SuperFlow

When to use SuperFlow

Mental model

The SuperFlow editor

Node library

Build with AI

Create an agent from the editor

Built for production workloads

Triggers

Expressions

Node types

AI

Apps

Control Flow

Safety

I/O and Compute

Data Transform

Human-in-the-Loop

Document

Utility

Running and monitoring

Sub-workflows and sub-agents

Next steps

​When to use SuperFlow

​Mental model

​The SuperFlow editor

​Node library

​Build with AI

​Create an agent from the editor

​Built for production workloads

​Triggers

​Expressions

​Node types

​AI

​Apps

​Control Flow

​Safety

​I/O and Compute

​Data Transform

​Human-in-the-Loop

​Document

​Utility

​Running and monitoring

​Sub-workflows and sub-agents

​Next steps

When to use SuperFlow

Mental model

The SuperFlow editor

Node library

Build with AI

Create an agent from the editor

Built for production workloads

Triggers

Expressions

Node types

AI

Apps

Control Flow

Safety

I/O and Compute

Data Transform

Human-in-the-Loop

Document

Utility

Running and monitoring

Sub-workflows and sub-agents

Next steps