Skip to main content

​
Understanding Agents

Agents are the intelligence layer of QuivaWorks. They’re AI-powered workers that can understand context, reason through problems, use tools, and make decisions—all within boundaries you define.

​
What is an Agent?

An agent is an AI that can:
  • Understand - Parse natural language and complex data
  • Reason - Think through problems step-by-step
  • Act - Use tools and APIs to accomplish tasks
  • Decide - Make judgements based on policies and context
  • Adapt - Handle exceptions and edge cases
Unlike rigid automation or simple chatbots, agents work more like human employees: you give them a goal, the tools they need, and boundaries to operate within—then they figure out how to accomplish the task. Agent Architecture

​
Agents vs. Automation vs. Chatbots

Traditional Automation

“If this, then that”Follows exact rules. Breaks on exceptions. Requires programming every scenario.❌ Can’t handle variability
❌ Needs explicit programming
❌ Brittle with edge cases

Simple Chatbots

“Conversation interface”Responds to keywords. Limited to chat. No real reasoning or actions.⚠️ Keyword-based only
⚠️ Can’t use tools
⚠️ Limited to conversation

quiva.ai Agents

“Digital workforce”Reasons through problems. Uses tools. Handles exceptions. Makes decisions.✅ Intelligent reasoning
✅ Uses tools dynamically
✅ Handles complexity
✅ Works within boundaries

​
Core Agent Capabilities

​
1. Intelligence & Reasoning

Agents can:
  • Understand natural language and intent
  • Reason through multi-step problems
  • Handle ambiguity and edge cases
  • Learn from context and conversation history
  • Apply business logic flexibly

Example: Customer Service

Customer: “I want to return this shirt but I lost the receipt. It’s been 35 days.”Traditional Automation: Breaks (doesn’t match exact return scenario)Agent:
  1. Understands customer wants to return an item
  2. Checks order history (using tool)
  3. Sees purchase was 35 days ago (5 days past standard policy)
  4. Reviews return policy (using tool)
  5. Notes customer has good history and high lifetime value
  6. Decides to approve as one-time exception
  7. Processes return and explains decision to customer

​
2. Tool Usage

Agents can use tools (connectors) to:
  • Search knowledge bases
  • Query databases
  • Call APIs
  • Access CRM data
  • Process documents
  • Send emails
  • Update systems
Agents decide which tools to use based on the task. You don’t program “if customer asks about order, call order API”—the agent figures that out.
Tools are MCP (Model Context Protocol) servers that provide standardized access to your systems. Learn more in Tools & Connectors.

​
3. Guardrails & Boundaries

Agents work within boundaries you define:
  • Output validation - Ensures responses match required format
  • Business rules - Enforces policies and constraints
  • Human-in-the-loop - Escalates critical decisions
  • Context limits - Controls memory and token usage
  • Response modes - Synchronous or background execution
Agents are powerful but not perfect. Always define boundaries, validate outputs, and use human-in-the-loop for critical decisions.

​
4. Self-Optimization

Agents improve automatically:
  • Smart Context - Intelligently manages conversation memory
  • Prompt Optimization - Automatically enhances prompts based on configuration
  • Adaptive Behavior - Learns from usage patterns (coming soon)

​
When to Use Agents

​
Perfect For

  • Answer customer questions with context
  • Troubleshoot issues using knowledge base
  • Apply policies with judgment (returns, refunds)
  • Handle complex, multi-turn conversations
  • Escalate to humans when needed
  • Ask discovery questions dynamically
  • Research companies and contacts
  • Score leads based on ICP criteria
  • Enrich data from multiple sources
  • Route qualified leads intelligently
  • Create personalized email campaigns
  • Generate social media posts
  • Adapt messaging by audience segment
  • Maintain brand voice across channels
  • Repurpose content efficiently
  • Extract information from documents
  • Validate data against business rules
  • Make contextual decisions on exceptions
  • Cross-reference multiple systems
  • Flag issues for human review
  • Personalize outreach at scale
  • Research prospects automatically
  • Follow up based on engagement
  • Book meetings intelligently
  • Qualify and route opportunities

​
Not Ideal For

When to Use Traditional Automation Instead

  • Simple, predictable tasks - If it’s always the same steps, use a condition or function
  • High-volume, low-variability - Agents have per-run costs; save them for complexity
  • Pure data transformation - Use Map, Rules, or Functions for straightforward data manipulation
  • Time-critical micro-operations - Agents add latency; use functions for speed-critical tasks
  • Deterministic calculations - Use Rules for exact math and logic
Rule of thumb: If you can write “if X then Y” rules that cover all cases, use automation. If there’s judgment, context, or exceptions, use agents.

​
Agent Architecture

Every agent has three main configuration areas:

​
1. Information Settings

Define your agent’s identity and behavior:
  • Name & Description - What this agent does
  • Response Mode - Wait for completion or run in background
  • Agent Instructions - Role, personality, and capabilities
Agent instructions are like a job description. Be specific about what the agent should do, how it should communicate, and what its limitations are.
Learn more about Information Settings →

​
2. Provider Settings

Configure the AI model and integration:
  • Provider - Workforce (custom), OpenAI, Claude, Gemini
  • Model - Specific model version (GPT-4, Claude Sonnet, etc.)
  • API Key - Securely stored authentication
  • Output Schema - Define required response structure
Different models excel at different tasks. GPT-4 for general intelligence, Claude for long context, Gemini for multimodal. Test to find what works best.
Learn more about Provider Settings →

​
3. Context Settings

Control memory and reasoning behavior:
  • Smart Context - Automatic conversation memory management
  • Prompt Optimization - Auto-enhance prompts
  • Token Limits - Maximum context size
  • Message History - How many messages to remember
  • Reasoning Steps - How many tool uses allowed
Higher token limits = more context = better decisions, but higher costs. Start with defaults and increase if needed.
Learn more about Context Settings →

​
Building Your First Agent

1

Define the Role

What is this agent responsible for? Customer service? Lead qualification? Content creation?
2

Write Instructions

Describe the agent’s role, personality, and capabilities clearly
3

Choose a Model

Select the LLM provider and model that fits your needs
4

Attach Tools

Connect the systems and data sources the agent needs
5

Set Boundaries

Define output schemas, validation rules, and escalation criteria
6

Test Thoroughly

Try edge cases, exceptions, and realistic scenarios
7

Deploy & Monitor

Activate in a flow and review performance regularly

Start Building

Follow our step-by-step guide to create your first agent in minutes

​
Agent Examples

​
Customer Service Agent

Role: Customer service representative for [Company]

Instructions: 
You help customers with order tracking, returns, and product questions. 
You're friendly, professional, and resolve issues on first contact.

Tools:
- Search Knowledge Base
- Get Order History
- Apply Return Policy
- Create Support Ticket

Boundaries:
- Approve returns up to $200 without escalation
- Escalate refunds over $200 to human
- Always verify customer identity before accessing orders

​
Lead Qualification Agent

Role: Sales development representative

Instructions:
You qualify inbound leads by understanding their needs, company fit, 
and timeline. You ask relevant discovery questions and score leads 
based on our ICP criteria.

Tools:
- Search CRM
- Company Data API (Clearbit/ZoomInfo)
- Calendar Booking
- Lead Scoring Rules

Boundaries:
- Only book meetings with qualified leads (score > 70)
- Maximum 5 discovery questions
- If lead doesn't match ICP, add to nurture campaign

​
Content Generation Agent

Role: Content marketing specialist

Instructions:
You create engaging social media posts and email campaigns that 
match our brand voice. You adapt messaging for different audience 
segments and channels while maintaining consistency.

Tools:
- Get Brand Guidelines
- Get Customer Segments
- Get Performance Metrics
- Content Calendar

Boundaries:
- Follow brand voice guidelines strictly
- Include call-to-action in every post
- Adapt length and tone by channel
- Flag sensitive topics for human review

​
Best Practices

Don’t say “help customers.” Say “help customers with order tracking, returns within policy, and product questions. Escalate refunds over $200.”
Connect data sources the agent needs. More context = better decisions. But don’t connect everything—focus on relevant tools.
Specify what the agent can and cannot do. Define escalation criteria. Use output schemas to enforce structure.
Use actual customer emails, edge cases, and exceptions. Agents learn from examples in instructions.
Begin with basic instructions and a few tools. Test thoroughly. Then add more capabilities and guardrails.
Review agent conversations regularly. Refine instructions based on performance. Update tools as systems change.
Add human approval for critical decisions, high-value actions, or when confidence is low.

​
Agent Costs

Agents have per-run costs based on:
  • Model chosen - GPT-4 > GPT-3.5, Claude Opus > Sonnet
  • Context size - More tokens = higher cost
  • Tool usage - Each tool call adds processing
  • Reasoning steps - More steps = more tokens
Use the Workforce model (our custom model) for the best balance of performance and cost. Or bring your own API keys to use any provider’s pricing.
See Plans & Pricing for details.

​
Next Steps

Create Your First Agent

Step-by-step guide to building an agent

Information Settings

Configure agent identity and behavior

Provider Settings

Choose models and configure AI

Context Settings

Manage memory and reasoning

Tools & Connectors

Connect your systems and data

Prompt Engineering

Write effective agent instructions

Best Practices

Patterns for successful agents

Use in Flows

Add agents to your workflows

​
Need Help?

Community

Ask questions and share agents

Support

Get help from our team

Examples

Browse agent templates