Creating Agents - Kore.ai Docs

Agents are the autonomous units of your application. Each agent has a role, an LLM for decision-making, tools to act with, and a prompt that defines its behavior.

Prerequisites

AgenticAI Core SDK installed and configured.
An LLM model and prompt configured. See Prompts and LLM Configuration.
Tools defined if the agent needs to perform actions. See Working with Tools.

Agent types

Autonomous agents

AI-powered agents that use LLM-based decision-making. Most agents use this type:

from agenticai_core.designtime.models.agent import Agent
from agenticai_core.designtime.models.llm_model import LlmModel, LlmModelConfig
from agenticai_core.designtime.models.prompt import Prompt

agent = Agent(
    name="FinanceAssist",
    description="Banking assistant for account management",
    role="WORKER",
    sub_type="REACT",
    type="AUTONOMOUS",
    llm_model=LlmModel(
        model="gpt-4o",
        provider="Open AI",
        connection_name="Default Connection",
        modelConfig=LlmModelConfig(
            temperature=0.7,
            max_tokens=1600
        )
    ),
    prompt=Prompt(
        system="You are a helpful banking assistant.",
        custom="Assist with account inquiries and transactions."
    )
)

Proxy agents

Agents that delegate processing to an external system:

agent = Agent(
    name="ExternalService",
    description="Integrates with external API",
    role="WORKER",
    sub_type="PROXY",
    type="AUTONOMOUS",
    llm_model=...,
    prompt=...
)

Agent sub-types

REACT (ReAct pattern)

The REACT sub-type uses a Reasoning + Acting loop. The agent iterates through:

Reason about the task.
Select an action (tool).
Observe the result.
Continue until complete.

Best suited for general-purpose tasks, interactive problem-solving, and tool-heavy workflows.

agent = Agent(
    name="ReactAgent",
    sub_type="REACT",
    ...
)

Agent roles

Role	Description
`WORKER`	Executes specific tasks delegated by the orchestrator.
`SUPERVISOR`	Coordinates other agents; typically used for the orchestrator agent.

# Worker
agent = Agent(
    name="BillingWorker",
    role="WORKER",
    description="Handles billing and payment tasks",
    ...
)

# Supervisor
agent = Agent(
    name="Supervisor",
    role="SUPERVISOR",
    description="Routes requests to appropriate workers",
    ...
)

Add tools to agents

Builder pattern

from agenticai_core.designtime.models.agent import AgentBuilder
from agenticai_core.designtime.models.tool import Tool

# Define tools
tool1 = Tool(name="GetBalance", type="toolLibrary", ...)
tool2 = Tool(name="Transfer", type="toolLibrary", ...)

# Build agent with tools
agent_config = AgentBuilder() \
    .set_name("BankingAgent") \
    .set_description("Banking operations agent") \
    .set_role("WORKER") \
    .set_sub_type("REACT") \
    .set_type("AUTONOMOUS") \
    .set_llm_model(llm_model) \
    .set_prompt(prompt) \
    .set_tool(tool1) \
    .set_tool(tool2) \
    .build()

agent = Agent(**agent_config)

Direct assignment

agent = Agent(
    name="BankingAgent",
    tools=[tool1, tool2, tool3],
    ...
)

Configure prompts

Basic prompt

prompt = Prompt(
    system="You are a helpful assistant.",
    custom="You help with banking operations."
)

With instructions

prompt = Prompt(
    system="You are a banking assistant.",
    custom="""You help customers manage accounts and transactions.

    ## Your Capabilities
    - Check account balances
    - Process transactions
    - Answer policy questions
    """,
    instructions=[
        "Never ask for passwords or PINs",
        "Always confirm transaction amounts",
        "Be professional and courteous"
    ]
)

With memory context

Use {{memory.store.field}} template variables to inject stored data directly into the prompt, reducing unnecessary tool calls:

prompt = Prompt(
    system="You are a helpful assistant.",
    custom="""You have access to customer information:

    <AccountInfo>
    {{memory.accountInfo.accounts}}
    </AccountInfo>

    Use this to provide quick responses without invoking tools.
    """
)

For full prompt and LLM configuration options, see Prompts and LLM Configuration.

Additional configuration

Agent icon

from agenticai_core.designtime.models.icon import Icon

agent = Agent(
    name="FinanceAssist",
    icon=Icon(
        name="avatar3",
        color="#D9D6FE",
        type="human"
    ),
    ...
)

Real-time (voice/audio)

agent = Agent(
    name="VoiceAgent",
    is_real_time_enabled=True,
    ...
)

Custom metadata

agent = Agent(
    name="CustomAgent",
    metadata={
        "versionInfo": "Draft version",
        "customField": "value"
    },
    ...
)

Register with the orchestrator

Convert an Agent to a lightweight AgentMeta for orchestrator registration:

# Convert full Agent to lightweight AgentMeta
agent_meta = agent.to_agent_meta()

# Use in orchestrator
app_agents = [agent.to_agent_meta() for agent in agents]
orchestrator = CustomOrchestrator(agents=app_agents)

Best practices

Naming: Use descriptive names that reflect the agent’s purpose. The orchestrator uses names for routing.
Descriptions: Write detailed capability descriptions so the orchestrator knows when to use the agent. Include known limitations and constraints.
LLM configuration: Match temperature to the task type. Set token limits appropriate to expected response length. Consider cost versus quality trade-offs.
Prompts: Be specific about the agent’s role. Include relevant context and clear guidelines. Use memory template variables to reduce unnecessary tool calls.
Tools: Only assign tools relevant to the agent’s scope. Avoid tool overload — too many tools increase latency and reduce routing precision.
Testing: Test agents individually before wiring them into a multi-agent system. Verify tool integration, memory access, and prompt effectiveness separately.

​Prerequisites

​Agent types

​Autonomous agents

​Proxy agents

​Agent sub-types

​REACT (ReAct pattern)

​Agent roles

​Add tools to agents

​Builder pattern

​Direct assignment

​Configure prompts

​Basic prompt

​With instructions

​With memory context

​Additional configuration

​Agent icon

​Real-time (voice/audio)

​Custom metadata

​Register with the orchestrator

​Best practices

​Related resources