Skip to main content

Creating Agents

This guide walks you through creating a new AI agent from scratch, covering configuration options and best practices.

Prerequisites

  • Access to Agent Studio
  • Understanding of your use case and goals
  • Knowledge of what tools the agent will need

Step 1: Start a New Agent

  1. Navigate to Agent Studio from the main navigation
  2. Click Agents in the sidebar
  3. Click Create Agent or the + button
You’ll see the agent configuration form.

Step 2: Basic Information

Agent Name

Choose a descriptive name that indicates the agent’s purpose:
Good NamesPoor Names
”Account Health Analyzer""Agent 1"
"QBR Prep Assistant""Helper"
"Renewal Risk Monitor""AI Bot”
Names should help users quickly understand what the agent does without reading the full description.

Step 3: Write Instructions

Instructions (the system prompt) define how your agent behaves. This is the most important configuration.

Structure Your Instructions

# Role
[Who the agent is and what it specializes in]

# Primary Tasks
[What the agent should do when asked]

# Guidelines
[How the agent should approach tasks]

# Output Format
[How responses should be structured]

# Constraints
[What the agent should avoid]

Example: Customer Health Agent

# Role
You are a customer success analyst specializing in account health assessment and early warning detection.

# Primary Tasks
When analyzing an account:
1. Review health score trends over the past 90 days
2. Analyze engagement patterns from activities
3. Identify key contacts and their recent involvement
4. Compare against similar successful accounts
5. Generate specific, actionable recommendations

# Guidelines
- Base all conclusions on data, not assumptions
- Prioritize actionable insights over general observations
- Consider both quantitative metrics and qualitative signals
- Explain your reasoning clearly

# Output Format
Structure your analysis as:

## Health Assessment
[Overall assessment with health score context]

## Key Findings
- [Finding 1 with data support]
- [Finding 2 with data support]

## Risk Factors
- [Risk 1]: [Mitigation suggestion]

## Recommended Actions
1. [Specific action with expected outcome]
2. [Specific action with expected outcome]

# Constraints
- Never fabricate data or statistics
- Don't make promises about outcomes
- If data is insufficient, say so explicitly

Step 4: Select the Model

Choose the AI model based on your needs:

Model Options

ModelSpeedCapabilityCostBest For
Gemini AutoVariesVariesOptimizedMost use cases
Gemini 3 ProSlowerHighestHigherComplex analysis, critical tasks
Gemini 3 FlashFastestGoodLowerQuick queries, simple tasks
Gemini 2.5 FlashFastVery GoodModerateBalanced needs

Recommendations

  • Start with Gemini Auto - Let the system optimize model selection
  • Use Pro for complex research, multi-step analysis, or critical decisions
  • Use Flash for quick lookups, simple formatting, or high-volume tasks

Step 5: Configure Reasoning Level

Reasoning level controls how much the agent “thinks” before responding:

Options

LevelBehaviorUse When
AutoAdapts to task complexityYou’re unsure what level is needed
LowQuick, direct responsesSimple questions, factual lookups
MediumModerate analysisStandard analysis, most tasks
HighDeep reasoningComplex problems, nuanced analysis
Higher reasoning = better analysis but slower responses and higher cost. Match the level to your actual needs.

Step 6: Enable Toolkits

Toolkits give your agent capabilities. Enable only what’s needed.

Available Toolkits

ToolkitCapabilitiesEnable When
Account DataGet account details, health scores, custom fieldsAgent needs account information
Activity DataFetch activities, emails, meetingsAgent analyzes engagement
Contact DataGet contact details, roles, engagementAgent works with people
Task ManagementCreate/update tasks, add commentsAgent manages follow-ups
Project DataGet project status, tasks, timelinesAgent reviews implementations
CommunicationSend emails, Slack messagesAgent takes outreach actions
CRM IntegrationSalesforce/HubSpot operationsAgent syncs with CRM
Knowledge BaseSearch documents, articlesAgent needs reference materials
Data AnalysisRun Python code for accurate counting, aggregation, and data calculationsAgent needs precise arithmetic or data analysis

Guidelines

Each enabled toolkit adds complexity. An agent with 3 focused toolkits will often outperform one with 10 toolkits.
A health analysis agent needs Account + Activity data. It probably doesn’t need Task Management or Communication.
Will users expect the agent to take actions (create tasks, send messages) or just provide information?

Step 7: Set Scope

Define what data the agent can access:

Scope Options

ScopeAccessUse For
GlobalAll organization dataGeneral-purpose agents
AccountSpecific account contextAccount-focused workflows

Account Scope

When using account scope:
  • Agent can only access data for the specified account
  • Useful for embedding agents in account-specific contexts
  • Provides data isolation and relevance

Step 8: Add Datasets (Optional)

Connect specific data sources for the agent to reference:
  • Knowledge bases - Internal documentation
  • Playbooks - Standard procedures
  • Templates - Response templates
This gives the agent access to your organization’s specific knowledge.

Step 9: Test Your Agent

Before deploying, test thoroughly.

Test Scenarios

  1. Typical request
    • Ask something your users commonly would
    • Verify the response is helpful and accurate
  2. Ambiguous request
    • Ask something vague
    • Does the agent ask clarifying questions or make reasonable assumptions?
  3. Edge case
    • Ask about an account with missing data
    • How does the agent handle incomplete information?
  4. Out of scope
    • Ask something outside the agent’s purpose
    • Does it gracefully redirect or explain limitations?

Example Test Prompts

Test 1: "Tell me about Acme Corp's health"
Expected: Detailed analysis with specific metrics

Test 2: "Acme Corp"
Expected: Asks what information is needed OR provides summary

Test 3: "What's the weather?"
Expected: Explains it can't help with that / redirects to purpose

Step 10: Deploy

Once tested:
  1. Click Save to save your agent
  2. Toggle Active to enable it
  3. The agent is now available for use

Configuration Examples

Research-Focused Agent

Name: Account Research Assistant
Model: Gemini 3 Pro
Reasoning: High
Toolkits: Account Data, Activity Data, Contact Data
Scope: Global

Instructions:
You conduct deep research on customer accounts. When asked about
an account, provide comprehensive analysis covering health metrics,
engagement trends, key contacts, and strategic recommendations.
Always cite specific data points and explain your reasoning.

Action-Oriented Agent

Name: Task Creator
Model: Gemini 2.5 Flash
Reasoning: Medium
Toolkits: Account Data, Task Management
Scope: Account

Instructions:
You help users quickly create and manage tasks. When asked to create
a task, gather the essential information (title, assignee, due date)
and create it immediately. Keep interactions concise and efficient.

Communication Agent

Name: Outreach Drafter
Model: Gemini Auto
Reasoning: Medium
Toolkits: Account Data, Activity Data, Contact Data
Scope: Account

Instructions:
You draft personalized outreach messages. When asked to draft an
email or message, research the account and contacts first, then
create tailored content that references specific context. Always
output in a ready-to-send format.

Best Practices Summary

Begin with minimal toolkits and instructions. Add complexity only when needed.
Vague instructions lead to inconsistent results. Be explicit about what you want.
Sample data doesn’t catch real-world issues. Test with actual accounts and scenarios.
Your first version won’t be perfect. Collect feedback and refine instructions.
Track how the agent is being used and whether responses are helpful.

Next Steps

Agent Best Practices

Advanced tips for agent optimization

Use Agents in Automations

Embed agents in workflow automations