AI Flows

AI Flows are the orchestration layer of Karyam.

While Agents focus on reasoning and decision making, AI Flows focus on execution and process orchestration.

Flows allow organizations to define:

• Business processes
• Multi-step automations
• Structured AI operations
• Approval workflows
• Human-in-the-loop interactions
• Reusable AI capabilities

If Agents are the brains of AI systems, AI Flows are the processes that connect business operations together.

---

Why AI Flows Exist

Many business operations require more than a single response.

Examples include:

• Employee onboarding
• VPN access requests
• Customer support escalation
• Expense approvals
• Incident management
• Content generation pipelines

These processes often involve:

• AI reasoning
• External systems
• Human approvals
• Multiple execution steps

AI Flows provide a structured way to coordinate these operations.

---

Creating an AI Flow

AI Flows can be created from:

BUILD
└── AI Flows
    └── Create Flow

Each flow consists of several configuration areas.

---

General Information

Name

The display name of the flow.

Examples:

• Employee Onboarding
• VPN Access Request
• Incident Triage
• AI Research Pipeline

---

Description

A short explanation describing the purpose of the flow.

Example:

Handles employee VPN access requests and approval routing.

---

Icon

An optional icon used to visually identify the flow within the platform.

Icons make large collections of flows easier to navigate.

---

Model & Tuning

Flows execute using a language model.

Model

Select the model responsible for executing the flow.

Examples include:

• GPT-4o
• Claude Sonnet
• Gemini
• Bedrock Models
• Ollama Models

---

Maximum Steps

Controls how many reasoning iterations the flow may perform during execution.

Range:

5 — 30

Higher values allow more complex workflows but may increase latency and cost.

---

Temperature

Controls creativity and response variability.

Value	Behavior
0.0	Deterministic
0.5	Balanced
1.0	Creative

---

Enable Caching

Caching allows supported providers to reuse previously generated results, reducing latency and cost.

Currently supported by:

• Anthropic
• Bedrock (Anthropic models only)
• MiniMax

Unsupported providers automatically ignore this setting.

---

Skills & Tools

AI Flows become useful when they can interact with external systems.

Skills

Flows can invoke Skills to interact with systems such as:

• Databases
• Email
• GitHub
• SSH
• Web Search

---

MCP Tools

Flows can invoke tools exposed through MCP Servers.

Examples include:

• Internal APIs
• Jira
• Slack
• Custom business systems

---

Sub-Flows

Flows can invoke other flows.

This enables organizations to build reusable building blocks.

Example:

Employee Onboarding
├── Create User Account
├── Provision Email
├── Create VPN Access
└── Notify Manager

---

Categories

Flows can be assigned to up to five categories.

Categories improve organization and discoverability within Service Catalogs.

Examples:

• Human Resources
• IT Operations
• Finance
• Customer Support

---

Inputs

Flows support dynamic input fields.

Input fields define variables that can be provided at execution time.

Examples:

employee_id
department
vpn_required
manager_email

Each input field can be referenced within prompts using placeholders.

Example:

${employee_id}
${department}
${manager_email}

---

System Instructions

System Instructions define the flow’s role, behavior, and operational boundaries.

These instructions execute before every flow run.

Example:

You are responsible for processing employee onboarding requests.

Rules:
- Verify required information exists.
- Request approval when required.
- Never provision accounts without authorization.

---

User Prompt

The User Prompt contains the primary instruction executed by the flow.

Unlike Agents, Flow prompts are often highly structured and parameterized.

Example:

Provision VPN access for employee ${employee_id}
working in department ${department}.

Placeholders must have corresponding input fields defined in the Inputs section.

---

Output Format

Flows can enforce a specific response structure.

Examples include:

JSON

{
  "approved": true,
  "ticket_id": "INC-12345"
}

Table

User	Status
John	Approved

Bullet List

• VPN Created
• Email Provisioned
• Manager Notified

If no output format is defined, the model returns its default response format.

---

API Access

AI Flows can be triggered programmatically using the Karyam API.

This allows flows to become part of existing business systems and automation pipelines.

---

Authentication

API requests use Bearer token authentication.

Authorization: Bearer <your_api_token>

Authentication can also be supplied using:

?token=<your_api_token>

---

Trigger Endpoint

Flows are executed using a POST request.

POST /api/v1/workspaces/{workspace_uuid}/flows/{flow_uuid}/execute

---

Synchronous Execution

Flows also expose a synchronous execution endpoint.

This endpoint waits for execution to complete before returning a response.

POST /api/v1/workspaces/{workspace_uuid}/flows/{flow_uuid}/execute/sync

Optional timeout support:

?timeout=60

Default timeout:

60 seconds

---

Request Body

Input placeholders can be provided as request parameters.

Example:

{
  "employee_id": "EMP-1001",
  "department": "Engineering",
  "vpn_required": true
}

If the flow contains no placeholders, the request body may be omitted entirely.

---

API Response

Submitting a flow returns execution metadata.

Typical response fields include:

• Job UUID
• Submission status
• Start time
• Execution message

The Job UUID can be used to monitor execution progress through Runs and Logs.

---

Relationship to Other Concepts

Models
    +
Skills
    +
MCP Servers
    +
Approvals
        ↓
AI Flow
        ↓
Runs

AI Flows coordinate business operations across multiple systems.

---

The Karyam Philosophy

Business processes require more than intelligence.

They require structure, approvals, observability, and governance.

AI Flows combine reasoning with operational control to create production-ready AI systems.

---

Agents decide. Flows execute. Businesses operate.