Create and manage custom AI assistants combining Knowledge Bases, Tools, and Guardrails
Tired of repeating context to the base AI model, manually pasting internal documents, and re-stating answer formats every time?An Agent bundles a Knowledge Base + Tool + Guardrail + System Prompt together to create a per-department-optimized AI assistant.
In one line — bundle a system prompt + knowledge base + tools + guardrails into a task-specific AI tailored per team.
When does it fit? When the same context repeats turn after turn, or when answer format and internal-source citations must stay consistent
How long to build? A minimal prompt-only agent takes ~5 min · a full setup with KB, tools, and guardrails is closer to 30 min
What about cost? Each response adds KB retrieval and tool routing on top of the base model, so token usage is higher — for casual questions, the base model is cheaper
Is this the right tool? If branching and loops dominate, look at Flows. If you only need one or two tools, Tools alone may be enough
The agent receives a user question and generates a response through this pipeline.A guardrail validates the input, the Knowledge Base retrieves related documents, tools (API, DB) are invoked when needed, and the LLM produces the final response.
Agent, Base Model, and Flow look similar but fit different situations.
Base Model
When — One-off questions, free-form answers, general knowledge.Casual prompts like “what does this code mean?” or “what’s today’s FX rate” are fastest and cheapest here. No KB retrieval or tool routing means lower token usage.
Agent
When — Repeated KB citations, fixed tone or format, guardrails required.HR policy Q&A, code reviewers, data analysts — the pattern of letting the LLM decide which KB to cite and which tool to call.
Flow
When — Branching, loops, or external API chains dominate.Form input → automated processing → notification, multi-step approvals, and similar automation. You specify the decision flow as nodes, so behavior is predictable.
When in doubt, start with an Agent. If it feels like overkill for simple questions, drop down to the base model; if you find yourself wanting if/else branches, move it to a Flow. You don’t have to pick the right one on the first try.
Click Workspace > Agents > ”+ New Agent” and fill in basic info.
Set the agent's name, description, and profile image
Field
Description
Example
Agent Name
Agent display name
”Marketing Assistant”
Description
What the agent does
”Marketing content creation and analysis support”
Agent Image
Agent icon
Marketing-related image
2
Pick the base model
Choose the AI model the agent will use. Pick from the model list registered by the admin.
3
Write the prompts
Define the agent’s role, persona, and response rules.
Field
Description
Task Prompt
Defines the agent’s role, persona, restrictions, and concrete task instructions. Plays the role of the general system prompt.
Response Format Prompt
Specifies the response format and structure (markdown, table, etc.). This is the default answer format — when Response Formats are registered below, it is used when no registered format fits the question.
Response Formats
Optional. Use ”+ Add Response Format” to register multiple answer styles per question type; the agent automatically picks the one that fits each question. Leave empty and the Response Format Prompt above always applies.
Click the Generate with AI button next to each prompt field — it analyzes the agent’s name, description, and connected resources to draft the prompt automatically.
During AI auto-generation, technical instructions (tool usage, SQL writing rules, etc.) are automatically excluded. The platform handles those — the prompt only contains role, persona, and restrictions.
Write the Task Prompt (role definition) and Response Format Prompt (output format) separately
Example of a good task prompt
You are Cloocus's marketing assistant.## Role- Support marketing content creation- Draft social media posts- Analyze marketing data## Response Rules- Always respond in Korean- Maintain a professional yet friendly tone- Provide data-driven insights- Comply with brand guidelines## Restrictions- Do not disparage competitors- Do not use unverified statistics
Why are the Task Prompt and Response Format Prompt separated?
The two prompts are used at different stages of agent execution.
Task Prompt
Response Format Prompt
When applied
While the agent is using tools
When composing the final answer
Role
”What to do” (role, restrictions)
“How to answer” (markdown, tables, length)
Include
Role definition, behavior rules, restrictions
Output format, tone, structure
Don’t include
Output format specs
Role definition, behavior rules
Separating them lets you change just the output format while keeping the role, or vice versa.
4
Configure prompt suggestions (optional)
Set conversation-starter suggestions shown when an agent is selected in chat.
Option
Description
Default
Use system default suggestions
Custom
Set agent-specific suggestions
Providing example questions matched to the agent’s purpose helps users start conversations quickly.
Examples: “Summarize this month’s sales”, “Draft a social media post”
5
Connect Knowledge Bases
Attach documents the agent should reference.
Click “Select Knowledge” in the “Knowledge Base” section
Select Knowledge Bases to connect (multiple supported)
Connected Knowledge Bases are searched via RAG and used in answers, with citations.
When you connect multiple Knowledge Bases, the agent automatically picks the right one based on the question
6
Connect databases (optional)
Connect a Database (DbSphere) for natural-language data queries (NL-to-SQL).
Click “Select Database” in the “Database” section
Select databases (multiple supported)
Ask questions in natural language about the connected DB; the AI generates and runs SQL, returning results.
7
Connect glossaries (optional)
Connect a Glossary so the agent understands your organization’s business terminology.
Click “Select Glossary” in the “Glossary” section
Select glossaries (multiple supported)
Term definitions, synonyms, and context registered in the glossary are reflected in agent responses.
8
Connect tools (optional)
Connect tools for external system integration. In the “Tool Connections” section, choose MCP servers or OpenAPI servers.
Tool Type
Description
OpenAPI server
Interact with external services via REST API
MCP server
Tool integration via Model Context Protocol
If a connected tool’s description is empty, a warning banner appears at the top of the editor. The tool description is the LLM’s primary cue for “when to call this tool”, so a missing description leads to incorrect calls or unused tools. When the banner appears, fill in the description from the tool definition.
9
Capability settings (optional)
Configure advanced features the agent can use. Each capability has 3 states.
State
Description
Disabled
The capability is completely hidden in chat (default)
Default On
Auto-enabled at chat start, user can turn off
Default Off
Visible in chat, but user must turn it on
Capability
Description
Web Search
Real-time web search for up-to-date info. Configurable result count and domain filter
Image Generation
AI image generation engine integration. Pick which connection to use
Code Interpreter
Run Python code for calculations and data analysis
Each capability is set to one of Disabled / Default On / Default Off
10
Response format (optional)
Constrain the agent’s response to a structured JSON format.
Mode
Description
Chat
Default freeform text response
Structured Output
Structured response per JSON Schema
In Structured mode, define the response schema with the visual field builder or the raw JSON editor.
Define the response structure with JSON Schema in Structured mode
11
Guardrail settings (optional)
Attach security guardrails to validate I/O.
Auto-detect and mask PII
Custom pattern filtering
Block prohibited words
LLM-based content validation
12
Auto-evaluation settings (optional)
Automatically monitor response quality.
Setting
Description
Judge Model
LLM to use for evaluation
Sampling rate
Share of responses to evaluate (1%–100%)
Evaluation type
Choose from retrieval quality, faithfulness, response quality
Evaluation type details
Type
Description
Retrieval Quality
Relevance of documents retrieved from the Knowledge Base
Faithfulness
Whether the response is faithful to retrieved content (no hallucination)
Response Quality
Overall quality, usefulness, and accuracy of the response
Sampling-rate guidance:
Situation
Recommended
Reason
New agent (validation phase)
50–100%
Need initial quality picture
Stabilized agent
5–10%
Save costs while monitoring
Critical-business agent
20–30%
Continuous quality assurance needed
Retrieval Quality and Faithfulness evaluations only run when there are Knowledge Base search results. For agents without a KB, choose Response Quality only.
Enable auto-evaluation to continuously monitor response quality
13
Access permissions
Set who can use the agent.
Option
Description
Public
Available to all users
Private
Available only to you
Group/Organization
Available to specified groups or organizations
Refine agent access by group or organizational unit
Concrete situations from operations and how to configure for them — use these as a starting point if your case is similar.
HR Assistant — Policy Q&A
Situation: HR fields the same questions about leave, benefits, and employment policy. Answers need to cite the specific policy article to be trusted.Configuration:
Field
Value
Base model
GPT-4o-mini (low cost, sufficient for fact-based answers)
Knowledge Base
HR policy, benefits guide PDFs
System prompt
”HR specialist. Always cite the article number at the end. Never speculate”
None (direct codebase access belongs in a separate MCP server)
Build sequence: Organize guidelines into a KB → create Agent → assign @code-reviewer handle → invoke in chat with @code-reviewer review this functionCommon pitfalls:
Build sequence: Create Agent → write system prompt only → paste notes in chatWhen this is enough: When you only want to bundle “prompt + model” without KB/tools/guardrails. 5 minutes to build. If even simpler, the Prompts feature may be all you need.
What's the difference between an agent and a base model?
Agents add Knowledge Bases, Tools, system prompts, and Guardrails to a base model, optimizing it for a specific task. Use the base model for general-purpose chat and agents for task-specialized chat.
Can I connect multiple Knowledge Bases to one agent?
Yes — connect multiple Knowledge Bases and databases at once. The agent automatically picks the appropriate resource based on the question. Writing a detailed Tool Description for each Knowledge Base improves selection accuracy.
Web Search / Image Generation / Code Interpreter doesn't work
Check the agent’s capability settings:
Disabled: The capability is completely hidden in chat
Default Off: User must turn it on in the chat input
Default On: Auto-enabled. If still not working, check admin settings (web search/image generation connections)
Code Interpreter requires both the agent setting and the user toggle in chat to be on.
Is agent usage tracked?
Yes — view per-agent usage, token consumption, and auto-evaluation results in the monitoring dashboard.
Can I move an agent to another environment?
Yes — use Export to download the JSON, then Import in the other environment. Note: connected Knowledge Bases, tools, and guardrails must be set up separately in the target environment.
Can I hide an agent?
Yes — use the toggle at the top-right of the agent card to change its state to Disabled, and it’s hidden from the chat model selector. The agent isn’t deleted and can be shown again.
