Agents
Agents are the primary intelligence layer of Karyam.
They are responsible for:
- Understanding requests
- Reasoning through problems
- Using tools and skills
- Interacting with external systems
- Collaborating with other agents
- Executing workflows
- Producing responses
If AI Flows define how work moves, Agents define how decisions are made.
Why Agents Exist
Section titled “Why Agents Exist”Traditional automation follows predefined rules.
Agents introduce reasoning and adaptability.
Instead of defining every possible path explicitly, agents can:
- Understand intent
- Decide which tools to use
- Retrieve knowledge
- Ask for clarification
- Delegate work
- Trigger workflows
This allows AI systems to operate in dynamic business environments.
Creating an Agent
Section titled “Creating an Agent”Agents can be created from:
BUILD└── Agents └── Create AgentEach agent consists of several configuration areas.
General Information
Section titled “General Information”Basic information used to identify the agent.
The display name of the agent.
Examples:
- IT Support Assistant
- HR Assistant
- Sales Copilot
- Procurement Agent
Description
Section titled “Description”A short explanation describing the purpose of the agent.
Example:
Handles employee VPN access requests and IT support tasks.Model & Tuning
Section titled “Model & Tuning”Agents require a language model for reasoning and execution.
Select the model that powers the agent.
Examples:
- GPT-4o
- Claude Sonnet
- Gemini
- Bedrock Models
- Ollama Models
Max Steps
Section titled “Max Steps”Controls how many reasoning steps an agent can perform during execution.
Range:
5 — 30Higher values allow more complex reasoning but may increase latency and cost.
Temperature
Section titled “Temperature”Controls creativity and response variability.
| Value | Behavior |
|---|---|
| 0.0 | Highly deterministic |
| 0.5 | Balanced |
| 1.0 | More creative |
Enable Caching
Section titled “Enable Caching”Caching can reduce latency and token consumption by reusing previously generated results.
Currently supported by:
- Anthropic
- Bedrock (Anthropic models only)
- MiniMax
Unsupported providers ignore this setting automatically.
Skills & Tools
Section titled “Skills & Tools”Agents become useful when they can perform actions.
Skills
Section titled “Skills”Skills provide access to external systems and capabilities.
Examples:
- PostgreSQL
- GitHub
- SSH
- Web Search
MCP Tools
Section titled “MCP Tools”Agents can invoke tools exposed by MCP servers.
Examples:
- Jira
- Slack
- Internal APIs
- Custom business systems
Sub-Agents
Section titled “Sub-Agents”Agents can delegate tasks to other specialized agents.
Examples:
Main Support Agent├── HR Agent├── Finance Agent└── IT AgentThis enables multi-agent systems.
AI Flows
Section titled “AI Flows”Agents can invoke AI Flows to execute structured business processes.
Examples:
- Employee onboarding
- VPN approval workflow
- Expense reimbursement
Chat Settings
Section titled “Chat Settings”Voice Input
Section titled “Voice Input”Agents can support voice conversations by attaching a Speech-to-Text skill.
This allows users to interact with agents using spoken language instead of text.
System Instructions
Section titled “System Instructions”System Instructions define the agent’s identity, responsibilities, constraints, and behavior.
They execute before every conversation and guide the model throughout the session.
Example:
You are an IT support assistant responsible for handling employee access requests.
Responsibilities:- Verify employee identity.- Retrieve company policies.- Request approvals when required.- Create support tickets.
Rules:- Never bypass approvals.- Always follow company policy.- Escalate uncertain situations to humans.System instructions are one of the most important parts of an agent’s design.
API Access
Section titled “API Access”Agents can be accessed programmatically through the Karyam API.
Authentication uses Bearer tokens.
Authorization: Bearer <your_api_token>Step 1 — Create a Session
Section titled “Step 1 — Create a Session”Create a session before sending messages.
POST /api/v1/workspaces/{workspace_uuid}/agents/{agent_uuid}/sessionsThe response contains a:
session_uuidThis session maintains conversation state and context.
Step 2 — Send Messages
Section titled “Step 2 — Send Messages”Use the returned session UUID to continue the conversation.
POST /api/v1/workspaces/{workspace_uuid}/agents/{agent_uuid}/sessions/{session_uuid}/chatExample request:
{ "message": "My VPN access has expired."}Agents can be embedded directly into websites using the Karyam Chat Widget.
This allows organizations to deploy AI assistants to customers, employees, or partners without building custom frontends.
Widget Features
Section titled “Widget Features”The embedded widget supports:
- Custom agent name
- Custom description
- Brand colors
- Domain restrictions
- Secure authentication
Allowed Domains
Section titled “Allowed Domains”Embedded agents can be restricted to specific domains.
Example:
company.comsupport.company.comportal.company.comOnly approved domains are allowed to load the widget.
Customization Options
Section titled “Customization Options”Available customization options include:
| Option | Purpose |
|---|---|
| agentName | Widget header title |
| agentDescription | Widget subtitle |
| primaryColor | Widget branding color |
Relationship to Other Concepts
Section titled “Relationship to Other Concepts”Models +Skills +MCP Servers +AI Flows ↓Agent ↓RunsAgents sit at the center of AI execution in Karyam.
They combine reasoning with actions.
The Karyam Philosophy
Section titled “The Karyam Philosophy”Agents should not simply answer questions.
They should understand context, make decisions, and perform work.
This is what transforms AI from a chatbot into a business system.