Build Your First Agent

A complete walkthrough to create, test, and deploy your first AI agent

This tutorial walks you through building a production-ready AI agent — from connecting knowledge to monitoring live conversations. Plan about 30 minutes.

If you have not completed the Quickstart yet, start there. This tutorial picks up where the quickstart leaves off and covers prompt engineering, tools, testing, and monitoring in depth.

Prerequisites

• A working AGO agent from the Quickstart (knowledge source connected, agent created, permissions set)
• Admin or staff permissions

---

1. Plan Your Agent's Scope

   Before refining your agent, define clear boundaries:

   Decision	Example
   Scope	Answer product questions for customers
   Tone	Professional, friendly, concise
   Escalation	Create a ticket when the agent cannot help
   Boundaries	No account modifications, no billing changes

   A focused agent performs much better than one that tries to handle everything. Start narrow, then expand.
2. Check Knowledge Quality

   Good answers depend on good source content.

   1. In the admin sidebar, click Knowledge, then select Quality
   2. Review the quality scores for your documents
   3. Fix any flagged issues (missing titles, thin content, outdated information)

   Spend time here — knowledge quality directly affects response accuracy.

   For details, see Knowledge Quality.

3. Write a System Prompt

   A good prompt is the single biggest lever for agent quality. Open your agent in the agent builder and go to the Prompt section.

   Example prompt:

   You are a friendly support assistant for ACME Corp.
   
   ## Your role
   - Answer product questions using the knowledge base
   - Guide users through common workflows
   - Create a support ticket when you cannot resolve an issue
   
   ## Rules
   - Always cite your sources when answering from documentation
   - If you are unsure, say so — do not make up answers
   - Keep responses concise (2-3 paragraphs maximum)
   - Use a professional but approachable tone
   
   ## Escalation
   Create a ticket when:
   - The user explicitly asks for human help
   - The issue requires account-level access
   - You have tried troubleshooting but cannot resolve the problem
   

   For writing techniques, see Prompt Engineering.

4. Add Tools

   Tools let your agent take actions beyond answering questions.

   1. In the agent builder, go to Tools
   2. Enable the Ticketing tool for human escalation
   3. Optionally enable:
      • HTTP Request — fetch data from external APIs
      • Send Email — send notifications
      • Parameterized Search — advanced knowledge retrieval

   For each tool, write a clear description of when the agent should use it. The agent relies on this description to decide.

   See Tools for all available options.

5. Test with Simulation

   Before deploying to real users, validate your agent with automated tests.

   1. In the admin sidebar, click AI Studio, then select Agent Lab
   2. Create a test dataset with sample questions and expected behaviors
   3. Run the simulation
   4. Review results — check for accuracy, tone, and correct tool usage
   5. Fix issues in your prompt or knowledge, then re-run

   Include edge cases — off-topic questions, adversarial inputs, and questions the agent should escalate.

   See Simulation Testing for the full testing workflow.

6. Deploy

   Choose how your users will interact with the agent:

   Channel	Documentation
   Widget	Widget Configuration
   Slack	Slack Integration
   API	API Reference
   SDK	SDK Integration

7. Monitor and Improve

   Your agent is live — now keep it performing well.

   1. Enable the Thread Evaluator — automatically scores every conversation for resolution quality. See Thread Evaluator.
   2. Review the Performance Dashboard — track resolution rates, satisfaction, and trends. See Performance Dashboard.
   3. Browse Conversations — spot issues by reading real interactions. See Conversations List.
   4. Act on Feedback — use satisfaction data to identify improvement areas. See Feedback Analytics.

   Set up a review cadence:

   Frequency	Action
   Daily	Check for flagged conversations
   Weekly	Review performance dashboard and CX score trends
   Monthly	Audit knowledge quality and update content

Checklist

• Agent scope and boundaries defined
• Knowledge quality reviewed and issues fixed
• System prompt written with clear rules and escalation criteria
• Ticketing tool enabled for escalation
• Simulation tests passing
• Deployed to at least one channel
• Thread evaluator enabled

---

What’s Next

• Prompt Engineering — advanced prompt writing techniques
• Agent Best Practices — design patterns and common pitfalls
• Router Agent — delegate to specialist agents
• Set Up Your Knowledge Base — deep dive into knowledge management