Core Concepts
Workflow
A sequence of connected nodes that automates one process. Workflows start with a trigger and include actions and logic nodes that execute in order. Each workflow keeps its own variables, execution history, and settings. You’ll find them under Workflows in the left sidebar. Example: A workflow that starts when a customer submits a form, uses AI to categorize the request, and routes it to the appropriate team.In the API and the MCP server, a workflow is called an application. Tools such as
applications_list and manage_application, and fields such as applicationId, all refer to workflows (applications). One application is one workflow — not a folder of them.Node
A single step in a workflow. Nodes are the building blocks you connect together. There are three types:- Triggers — Start the workflow
- Actions — Perform tasks
- Logic — Control the flow
Execution
A single run of a workflow. When a trigger fires, it creates an execution (shown as a run in the app) that moves through each step in sequence. Open Runs to see what happened.Run per event / One shared run
How a workflow’s triggers start runs, set under How triggers start runs in Settings. A run per event starts a fresh run for every trigger event; One shared run sends every event into the same long-running run. The API and older docs call them Multi instance and Single instance. See Execution modes.Version
A saved copy of a workflow’s steps, connections and variables, made automatically as the workflow changes. Browse and restore them from Version history. See Version history.Trigger Types
Start Trigger
The simplest trigger. Runs the workflow when you click the Run button manually. Good for testing or workflows you want to run on-demand.Scheduled Trigger
Runs the workflow on a time-based schedule. Can be set to hourly, daily, weekly, or custom intervals using cron expressions. Example: Run a report every Monday at 9 AM.App Trigger
Starts the workflow when something happens in a connected application. Requires an integration to be set up first. Example: When a new row is added to a Google Sheet, when an email arrives in Gmail.Webhook
Starts the workflow when an external system sends an HTTP request to a unique URL. Useful for connecting with systems that support webhooks. It can also send the workflow’s result back to the caller, or a link to fetch it later. See Webhook. Example: When Stripe sends a payment notification, when a form service submits data.AI Trigger
Uses AI to monitor a condition and triggers when that condition is met. More flexible than rule-based triggers. Example: Trigger when an email appears urgent based on its content, not just keywords.Action Types
LLM
Sends a prompt to an AI model and returns the response. Use this for text generation, summarization, classification, or any task where you need AI intelligence.Integration Action
Performs an action in a connected application. Requires an integration to be set up. Example: Send an email via Gmail, create a row in Google Sheets, post a message to Slack.Execute Workflow
Runs another workflow you’ve built and returns its output to the calling workflow. Use it to compose workflows — build small, focused workflows and reuse them as building blocks.Execute Code
Runs custom Python code, for logic that isn’t available through standard nodes. The code can also call the node’s connected apps as functions — see Execute Code integrations.HTTP Request
Makes a web request to any URL. Use this to integrate with APIs that don’t have a dedicated CogniAgent integration.Update Variable
Sets or updates a variable value that persists for the current run or across every run of the workflow.Read File
Extracts content from uploaded files. Supports PDFs, documents, spreadsheets, and images.Logic Types
If / else
Branches the workflow based on a true/false condition. If the condition is true, one path executes. If false, another path executes. Example: If order total is greater than $100, apply free shipping. Otherwise, calculate shipping cost.Switch
Branches the workflow based on multiple conditions. Like a condition node but with more than two paths. Example: Route support tickets to different teams based on category: billing, technical, or general.Loop
Repeats a set of nodes for each item in a list. Useful for processing multiple records. Example: For each row in a spreadsheet, send a personalized email.Pause
Stops the workflow for a specified duration before continuing. Example: Wait 24 hours before sending a follow-up email.Wait and Combine
Waits for multiple parallel branches to complete before continuing. Used when you’ve split a workflow and need to merge the results.Cowork Concepts
Coworker
An AI colleague you hire, give a job to, and hand work to. Unlike a workflow from Workflows, you give it a goal rather than a sequence of steps — it works out how to get there and adjusts as it goes. See AI coworkers that do the work. Example: A research coworker that answers “how do our three biggest competitors price?” with a cited brief.Task
One job given to a coworker. A task has its own conversation and its own files, and is private to whoever started it. See Managing your tasks.Capability
Something a coworker is allowed to do — search the web, run code, drive a browser, talk to people, use a connected app. Each is a switch you control. See What a coworker can do.Skill
A procedure written down once so a coworker follows it the same way every time. A capability is something it can do; a skill is knowing how and when to do it. See What skills are.Delegation
One coworker handing part of a job to another, the way a team does. Also called a handover. The other coworker takes it on as a new task of its own and reports back when it finishes. It works with its own instructions and connected accounts. The capability is off by default, and every handover waits for your approval until you choose Always allow. See Working with coworkers, and Coworker handovers in detail for the actions and rules behind one.Dynamic workflow
A coworker’s plan for one large job: a team of temporary helpers working side by side, then one combined answer. Make repeatable on a completed run saves it, so you can re-run it by name from the slash menu. It also adds a workflow under Workflows that runs it. The dynamic workflow itself is not a workflow from Workflows. See Dynamic workflows, and Dynamic workflow scripts for the script behind one.Permission rule
A standing decision about what a coworker may always do, must always ask about, or may never do. When more than one rule matches, the strictest wins. See Permission rules.Files
Workspace files
The one folder a whole workspace shares. People open it at Cowork → Files, coworkers with Code and files reach it atfiles/ in their tasks, workflows read and write it, and a chat message names a file in it as @files/ and its path. Every change keeps its history. See Workspace files.
Checkpoint
A version of a file in Workspace files kept under a name until someone removes it, outside the file’s last three versions. A file keeps up to three. See Editing and file history.Conflicted copy
What’s saved beside a file when two changes to it crossed:report (conflicted copy).md, so neither change is lost. Compare it with the original, then keep one. See Editing and file history.
Workspace mount
A file or folder from Workspace files that an Execute Code step copies into the code’s working folder, underfiles/. Read-only mounts are never saved back; read-and-write ones are saved as new versions when the code succeeds. See Use a SQLite database with Workspace files.
Data Concepts
Variable
A named value that stores data during workflow execution. Variables can be set by nodes and referenced by other nodes. Types:- Runtime variables — Exist only during a single execution
- Workflow variables — Persist across executions. In the API and MCP tools these are workflow (application) variables, as in
create_application_variable.
Variable Reference
The syntax for accessing data from other nodes:{{node_name.field_name}}
Example: {{ask_ai_1.response}} gets the response from an LLM node named “ask_ai_1”.
Integration Concepts
Integration
A connection to an external service (Gmail, Slack, Shopify, etc.). Integrations require authentication and enable App Triggers and Integration Actions. CogniAgent supports 2,700+ integrations through Pipedream and built-in connectors. Learn more →Pipedream
CogniAgent’s primary integration platform, providing access to 2,700+ pre-built app connections. Handles authentication, rate limiting, and API updates automatically.MCP (Model Context Protocol)
A standard protocol for connecting AI models to external tools and data sources. CogniAgent both uses MCP (agents can call MCP-exposed tools) and exposes an MCP server so AI assistants like Claude can build and operate your workflows (applications) and Chat & Voice Agents (conversation flows) on your behalf. See the MCP server guide.AI Concepts
AI Agent
An autonomous AI system that can perform multi-step tasks, use tools, and make decisions. Unlike simple AI prompts, agents can take actions and adapt based on results.Human-in-the-Loop
A workflow pattern where automation pauses to get human input or approval before continuing. Ensures humans stay in control of important decisions. Example: AI drafts a response, but a human reviews and approves before sending. In CogniAgent, the Human Step node puts this pattern into a workflow.Prompt
The text instruction sent to an AI model. Good prompts are clear, specific, and include relevant context.Chat & Voice Agent Concepts
Chat & Voice Agent
An AI agent that talks to people — on your website, over the phone, and across Slack, WhatsApp, SMS and email. You build it on a visual canvas as a flow: one Start node, one or more actors, and routes between them. You’ll find them under Chat & Voice Agents in the left sidebar.In the API and the MCP server, a Chat & Voice Agent is called a conversation flow. Tools such as
list_conversation_flows and manage_conversation_flow refer to Chat & Voice Agents (conversation flows). The editor still says “flow” in places — Test flow, Agent settings, Untitled Flow — meaning the agent’s canvas.Actor
One of the AI roles inside a Chat & Voice Agent. Each actor has a name, an icon, a system prompt (context + instructions), a model, capabilities (tools, knowledge bases, integrations, hooks), optional steps, and a focus mode. You edit it in the actor’s panel, one section at a time: Instructions, Steps, Tools, Knowledge, Integrations and Hooks. Learn more →Start Node
The single entry point of an agent’s flow. Owns the channel configuration (Widget, Gmail, Slack, Teams, Phone). Every flow has exactly one.Route
A connection between two actors, also called a routing edge (the API and MCP tools call it an edge). Each route carries an activation prompt — a plain-language description of when this route should fire. The platform reads it during routing decisions and auto-generates a short label. A route into an actor can also carry an Only when… condition, which keeps the router from choosing that actor until a parameter or slot matches, or until an AI judges that a condition you wrote is met (Conditional routes). Learn more →Global Context
Shared knowledge written once on the flow and inherited by every actor that opts in. Use it to write your company introduction or product blurb once, not on every actor. Learn more →Focus Mode
How strictly an actor sticks to its job when users change topics mid-conversation. One of Auto, Flexible, Persistent, or Strict. Learn more →Definition of Done
A short rule on an actor describing when its job is complete. Triggers a handoff to the next actor.Slot
A value an agent collects during a conversation, such as an order number. An Ask Question step saves it under its Slot name, and prompts and conditions read it as@slot.<name>. A slot lasts for the whole conversation and travels with it from actor to actor.
Learn more →