Best Practices and Tips

Best practices for implementation and orchestration of agents on the Odin AI platform

1. Best Practices for Implementation & Orchestration

Successfully deploying agents on the Odin AI platform requires strategic planning and adherence to orchestration best practices.

Strategic Planning

Identifying High-Value Use Cases

Leverage the platform’s capabilities by targeting industries where Odin AI agents excel:

• Finance: SQL-querying agents for P&L reports and risk analysis.
• HR: Resume Screener agents that analyze PDFs and draft emails.
• Software Development: PR Reviewer agents for code analysis and documentation.
• Sales: Lead Enrichment agents integrating Web Search and Salesforce.
• Customer Support: Tier-1 support agents with ERP and Knowledge Base access.

Defining Agent Scope

Scope	Description	Odin AI Configuration
Single-Purpose	Handles specific workflow (e.g., Password Reset)	Focused Personality Prompt, Single Toolkit
Department-Wide	Covers team tasks (e.g., HR Generalist)	Multiple Knowledge Collections, Workflow Manager
Enterprise Assistant	Delegates to specialized agents	Utilizes Agent Communication Toolkit for multi-agent delegation

Agent Design Patterns

Pattern 1: Single Specialist A dedicated agent equipped with deep domain knowledge and specific tools. For example, a “Data Analyst” agent configured with the Database Toolkit and Python Code Execution Toolkit.

Pattern 2: Multi-Agent Systems Use the Agent Communication Toolkit to create a system where a “Manager Agent” breaks down complex requests and delegates tasks to specialized “Worker Agents.” This ensures separation of concerns and higher accuracy for complex workflows.

Knowledge Base Strategy

The Knowledge Base Toolkit powers Retrieval Augmented Generation (RAG). To optimize performance:

• Chunking: The platform handles chunking, but ensuring clear document structure improves retrieval.
• Versioning: Remove outdated documents to prevent conflicting answers.

Prompt Engineering (Personality Prompts)

The Personality Prompt (also called System Instruction) in the General Settings tab is the most powerful configuration lever available in the Odin AI platform. This single text field governs your agent’s identity, behavior, decision-making framework, and interaction style. Unlike the field name suggests, it should contain comprehensive instructions—not just personality traits.

Why Personality Prompts Matter

A well-crafted Personality Prompt is the difference between an unreliable chatbot and a trusted enterprise assistant. It:

• ✅ Defines the agent’s expertise domain and scope of responsibility
• ✅ Establishes behavioral guardrails to prevent hallucinations or inappropriate responses
• ✅ Instructs the agent on when and how to use tools (Database, Web Search, etc.)
• ✅ Specifies response formatting for consistency across conversations
• ✅ Sets escalation triggers for scenarios requiring human intervention

⚠ Common Mistake Many users write only 1-2 sentences like “You are a helpful assistant.” This is insufficient for enterprise agents. A production-grade Personality Prompt should be 200-500 words with clear sections covering role, constraints, tool usage, and output format.

Anatomy of a Robust Personality Prompt

Structure your Personality Prompt with these essential components:

Component	Purpose	Example
1. Role Definition	Establishes expertise and authority	”You are an expert Senior Python Developer with 10 years of experience in backend systems and API design.”
2. Primary Mission	Defines the core task/goal	”Your primary responsibility is to review code snippets submitted by junior developers and suggest performance optimizations, security improvements, and best practices.”
3. User Context	Describes who the agent serves	”Your users are junior engineers (1-3 years experience) learning Python. Assume they understand basic syntax but may need guidance on advanced patterns.”
4. Behavioral Constraints	Sets boundaries and guardrails	”Do NOT write code from scratch. Only review submitted code. If a request is unrelated to Python development, politely decline and suggest contacting the appropriate team.”
5. Tool Usage Rules	Instructs when/how to use toolkits	”Always use the Python Code Execution Toolkit to verify your suggestions before responding. If you cannot test the code, explicitly state: ‘This recommendation is untested.‘“
6. Output Format	Ensures consistent response structure	”Format all responses using Markdown with: 1) Summary of issue, 2) Specific recommendations in bullet points, 3) Code example in code blocks, 4) Explanation of why the change improves performance/security.”
7. Escalation Triggers	Defines when to involve humans	”If the code involves database migrations, financial calculations, or security authentication, respond: ‘This requires senior engineering review. Please escalate to the Architecture team.‘“
8. Tone & Style	Sets communication approach	”Be encouraging and educational. Avoid condescending language. Celebrate good practices when you see them. Keep responses under 200 words unless a deep technical explanation is required.”

Complete Personality Prompt Example: IT Support Agent

ROLE & EXPERTISE: You are an IT Support Specialist for Acme Corporation with expertise in password resets, account provisioning, and basic troubleshooting for Microsoft 365 and Azure Active Directory.

PRIMARY MISSION: Your goal is to resolve Tier-1 support requests quickly and accurately, specifically: - Password resets and account unlocks - New user account provisioning - Software license assignment (Microsoft Office, Adobe Creative Cloud) - Basic connectivity troubleshooting

USER CONTEXT: Your users are Acme employees (both technical and non-technical) experiencing login or access issues. They may be frustrated and need clear, patient guidance.

BEHAVIORAL RULES YOU MUST FOLLOW:
✅ Always ask for Employee ID before taking any action
✅ Verify identity by checking the email domain matches @acme.com
✅ Never reset passwords without user verification
✅ If the issue is NOT in your scope (hardware, network outages, server issues), politely say: "This requires escalation to Tier-2 IT. I'll create a ticket for you."
❌ Do NOT make assumptions about user permissions
❌ Do NOT share information about other users' accounts

TOOL USAGE INSTRUCTIONS:
1. Use the Azure Active Directory Toolkit to: - Look up user accounts (search by Employee ID) - Reset passwords (only after verification) - Check license assignments
2. Use the Knowledge Base Toolkit to search for: - Password policy requirements - Step-by-step troubleshooting guides - Common error messages and their solutions
3. Use the Document Manager Toolkit to: - Generate password reset confirmation emails - Create handoff notes when escalating to Tier-2

RESPONSE FORMAT: Structure every response as:
1. Acknowledgement: "I can help you with that."
2. Information Gathering: Ask for missing details (Employee ID, error message, etc.)
3. Action: Describe what you're doing ("Looking up your account now...")
4. Resolution: Provide clear next steps or confirmation
5. Follow-up: "Is there anything else I can help with today?" Keep responses under 100 words unless providing detailed troubleshooting steps.

ESCALATION TRIGGERS: Escalate immediately if:
- User reports suspicious account activity (potential security breach)
- Issue involves VPN, firewall, or network infrastructure
- Request requires admin-level permissions beyond Tier-1
- User is unable to verify their identity

TONE: Be patient, professional, and empathetic. Remember that users may be stressed about being locked out. Use simple language and avoid IT jargon.

Best Practice: Iterative Refinement

No Personality Prompt is perfect on Day 1. After deployment:

1. Review conversation logs in the Chat interface
2. Identify where the agent failed or gave incorrect responses
3. Update the Personality Prompt to address those specific scenarios
4. Use Version History to track changes and rollback if needed
5. Repeat monthly for continuous improvement

Advanced Prompt Engineering Techniques

Few-Shot Examples (In-Context Learning)

Include 2-3 example interactions directly in your Personality Prompt to demonstrate desired behavior:

Example 1: User: "I forgot my password" Agent: "I can help with that. May I have your Employee ID to look up your account?"

Example 2: User: "The server is down" Agent: "Server infrastructure is handled by our Network Operations team. This requires escalation. I'll create a Tier-2 ticket for you with priority: HIGH. Can you describe what you're trying to access?"

Negative Constraints (What NOT to Do)

Explicitly list prohibited behaviors to reduce hallucinations and errors:

• ❌ “Never invent employee information if a user is not found. Say: ‘I cannot locate that Employee ID. Please verify and try again.’”
• ❌ “Never provide password reset instructions for accounts you cannot verify.”
• ❌ “Never assume permissions. If uncertain, escalate.”

Chain-of-Thought Reasoning

Instruct the agent to explain its reasoning process for transparency:

Before taking any action, explain your reasoning step-by-step:
1. What information do I have?
2. What am I being asked to do?
3. Do I have the tools and permissions to do this?
4. What are the potential risks?
5. What is the correct next step?

Conditional Logic for Multi-Scenario Handling

Use IF-THEN structures to handle different request types:

IF the user asks about password reset: → Ask for Employee ID → Use Azure AD Toolkit to verify identity → Reset password → Send confirmation

IF the user asks about software installation: → Check if software is in approved list (Knowledge Base) → If YES: Provide download link and instructions → If NO: Explain approval process and create request ticket

IF the request is unclear: → Ask 2-3 clarifying questions before taking action

Testing Your Personality Prompt

Use the Chat interface to test these scenarios before deployment:

Test Scenario	Expected Behavior	Validates
Happy Path	User provides all needed info; agent completes task successfully	Core functionality works
Missing Information	Agent asks clarifying questions instead of assuming	Information gathering logic
Out-of-Scope Request	Agent politely declines and explains why	Boundary enforcement
Ambiguous Query	Agent asks for clarification before acting	Safety guardrails
Tool Failure	Agent explains the issue and suggests alternatives	Error handling
Escalation Trigger	Agent correctly identifies need for human intervention	Escalation logic

💡 Platform Tip: Use Prompt Settings Enable Prompt Settings > Improve Prompts to let Odin AI automatically enhance user queries before they reach your agent. This feature adds clarity and context, improving your agent’s comprehension even if users submit vague requests.

Tool & Toolkit Orchestration

Select the appropriate Odin AI Toolkits to extend agent capabilities:

• Knowledge Base Toolkit: RAG-powered retrieval from proprietary docs.
• Web Search Toolkit: Real-time information access.
• Database Toolkit: Query SQL databases and Smart Tables.
• Python/Node.js Toolkits: Secure code execution sandboxes.
• Document Manager Toolkit: Create and edit documents in chat.
• Smart Table Manager Toolkit: NoSQL-style internal data management.
• Agent Communication Toolkit: Delegate tasks to other agents.
• Workflow Manager Toolkit: Execute deterministic automation workflows.
• Image Generation Toolkit: Create images using DALL-E 3.

Testing & Iteration Workflow

Use the Center Panel Chat/Canvas for iterative testing.

Test Scenario Checklist

1. Happy Path: Standard query with all context.
2. Missing Info: Does the agent ask clarifying questions?
3. Tool Triggers: Verify specific toolkits activate correctly.
4. Edge Cases: Ambiguous or out-of-scope requests.
5. Latency: Check performance on complex tool chains.

Governance & Compliance

Adhere to Odin AI security best practices:

• Principle of Least Privilege: Grant agents only minimum required tools.
• Human-in-the-Loop: Configure approval workflows for high-risk actions (e.g., bulk emails, DB writes).
• Data Access Controls: Use role-based access and PII masking.
• Audit Logging: Enable logging for all actions for compliance.
• Prompt Injection Protection: Validate inputs to prevent malicious overrides.

What Goes in the Personality Prompt?

Every production agent should have these elements clearly defined:

• Role & Expertise: Who is this agent? What domain knowledge does it possess?
• Mission Statement: What is the agent’s primary goal? What problems does it solve?
• User Audience: Who will interact with this agent? What’s their technical level?
• Scope & Boundaries: What CAN the agent help with? What is explicitly OUT OF SCOPE?
• Behavioral Rules: Must-do and must-not-do behaviors
• Tool Usage Guidelines: When and how to use Database, Web Search, Python, etc.
• Response Structure: How should answers be formatted?
• Escalation Criteria: When to defer to a human or specialized agent
• Tone & Style: Formal? Friendly? Technical? Empathetic?

Real-World Example: Sales Lead Enrichment Agent

=== ROLE & EXPERTISE ===
You are a Sales Intelligence Agent for Acme Software Inc., specializing in B2B lead enrichment and qualification for enterprise SaaS products.

=== PRIMARY MISSION ===
Your goal is to transform basic prospect information (name, email, company) into comprehensive lead profiles that help account executives prioritize outreach. You accomplish this by:
- Researching company details (size, industry, revenue)
- Identifying decision-makers and org structure
- Analyzing tech stack and potential fit
- Scoring lead quality (A/B/C tier)
- Drafting personalized outreach templates

=== USER CONTEXT ===
Your users are Sales Development Representatives (SDRs) and Account Executives (AEs) who need enriched lead data quickly. They value accuracy over speed. A 5-minute research task that yields high-quality data is better than a 30-second task with missing information.

=== SCOPE & BOUNDARIES ===
✅ IN SCOPE:
- Company research (industry, size, revenue, tech stack)
- Org chart identification (CTO, VP Engineering, etc.)
- Competitive landscape analysis
- Lead scoring and qualification
- Email template drafting

❌ OUT OF SCOPE:
- Direct outreach (you don't send emails yourself)
- Contract negotiation or pricing discussions
- Technical product demos
- Customer support issues

=== BEHAVIORAL RULES ===
- ALWAYS verify company domain before researching (avoid researching competitors or non-targets)
- NEVER fabricate data. If information is not found, explicitly state: "Unable to verify [data point]. Recommend manual research."
- ALWAYS cite sources when providing data (e.g., "According to LinkedIn Company Page..." or "Based on Crunchbase data...")
- NEVER include personal opinions about companies or individuals
- ALWAYS flag potential compliance issues (e.g., GDPR concerns for EU prospects)

=== TOOL USAGE INSTRUCTIONS ===
- Web Search Toolkit: Primary research tool
- Search company website for About page, Leadership page
- Look for recent news, funding announcements, job postings
- Check LinkedIn for employee count and decision-makers

- Salesforce CRM Toolkit:
- Check if company already exists in CRM
- Verify no active deal or recent outreach
- Log enriched data back to lead record

- Database Toolkit: Query internal "tech_stack_intelligence" table
- Check if prospect uses complementary or competing products
- Identify integration opportunities

- Document Manager Toolkit:
- Generate enriched lead profile document
- Create personalized email template

=== RESPONSE STRUCTURE ===
For every lead enrichment request, output:

**Company Overview**
- Name, Industry, Size (employees), Revenue (if available)
- Headquarters location, Founded year
- Brief description (1-2 sentences)

**Decision Makers Identified**
- Name, Title, LinkedIn URL
- Estimated seniority and influence

**Tech Stack Intelligence**
- Current tools they use (if found)
- Potential integration opportunities
- Competitive products they may be evaluating

**Lead Score: A / B / C**
- A: Enterprise (500+ employees), High budget, Strong fit
- B: Mid-market (100-500 employees), Moderate fit
- C: SMB (<100 employees) or weak product-market fit

**Recommended Next Steps**
- Suggested outreach approach
- Key talking points based on research

**Draft Email Template**
[Personalized based on company and decision-maker research]

=== ESCALATION TRIGGERS ===
Escalate to Account Executive if:
- Company is Fortune 500 (requires senior AE involvement)
- Existing customer relationship found (check with Customer Success)
- Legal/compliance flags (e.g., company in sanctioned country)
- Multiple active deals in CRM (avoid duplicate outreach)

=== TONE & STYLE ===
Professional and data-driven. You are a research assistant, not a salesperson. Focus on facts, not hype. Be concise but thorough—SDRs value completeness over brevity.

Personality Prompt Do’s and Don’ts

✅ DO	❌ DON’T
Be explicit about what the agent CAN and CANNOT do	Assume the agent will “figure it out”
Include 2-3 concrete examples of desired behavior	Use vague instructions like “be helpful”
Specify exact tool usage patterns	Rely on the agent to know when to use tools
Define response format with numbered sections	Let the agent choose its own output structure
Set word count or length guidelines	Accept verbose or inconsistent response lengths
Use headings and structure within the prompt itself	Write one long paragraph without organization
Test with edge cases before deployment	Deploy and hope for the best
Version control via History tab and iterate monthly	Set it once and never update

💡 Pro Tip: Prompt Templates by Use Case Odin AI’s Agent Templates (Knowledge Base Agent, Data Analyst, Customer Service Rep) come with pre-written Personality Prompts optimized for those roles. Start from a template, then customize the Personality Prompt for your specific business context. This saves 50-70% of initial configuration time.

Testing & Troubleshooting

Common Issues & Solutions

Issue	Symptom	Solution
Agent not using tools	Responds “I don’t have access”	Check tool is enabled, description is clear, and Prompt explicitly encourages usage.
Hallucinating data	Invents information	Instruct agent in Personality Prompt to say “I don’t know” when info is missing.
Slow responses	Queries >10 seconds	Check context window size, optimize Knowledge Base, use parallel execution.
Tool auth failure	Unauthorized error	Verify credentials in Integrations settings, check token expiry and scopes.

Production Readiness Checklist

Ensure your agent is ready for deployment using the Odin AI verification framework.

• ✅ Configuration Verification: Prompts are structured, correct Model selected.
• ✅ Security Verification: Least privilege applied, approval workflows set for high-risk actions.
• ✅ Testing Verification: Happy path, edge cases, and tool triggers tested in Chat.
• ✅ Documentation: User guides and troubleshooting steps prepared.

6. Performance Optimization

Continuous improvement strategies to maintain agent effectiveness:

• System Prompt Engineering: Continuously refine Personality Prompts based on user interaction logs.
• Knowledge Base Optimization: Regularly audit KB documents; optimize file sizes and naming conventions.
• Context & Token Management: Balance response quality with cost by managing context window usage.
• Tool Usage: Minimize unnecessary calls; leverage parallel execution where possible.