Skip to content
Module 11The Complete Agent Lifecycle — End to End

Goal: Assemble an evidence-based rollout checklist.

Prerequisites: Bring your brief, instructions, approved knowledge, and test notes from the earlier lessons.

Expected result: A list of checked behaviors, unresolved failures, and the person responsible for each follow-up.

Bringing It All Together

You have now learned every component of the raia platform. This final lesson walks through the complete lifecycle of building and deploying an agent from start to finish, connecting all the modules together.

Phase 1: Plan (Module 2)

Before touching the platform:

  1. Answer the 5 Planning Questions (objective, audience, tone, guardrails, escalation rules)
  2. Document your agent's "North Star" — the specific outcome it should achieve
  3. Define 3–5 critical negative constraints
  4. Identify your escalation triggers (knowledge gaps, complexity, sentiment)

Phase 2: Build (Modules 3–5)

In raia CX Launch Pad:

  1. Info tab: Set agent name, description, and avatar
Agent identity section tab.
  1. Highlight 1: The Info tab.
Public and internal naming fields.
  1. Highlight 1: Public Name: the name people see when they talk to the agent.
  2. Highlight 2: Internal Name (make it unique).
Description and grouping fields.
  1. Highlight 1: Description, with Optimize with AI.
  2. Highlight 2: Tags: labels that help your team group agents.
  1. Train tab: Write system instructions, optionally using the suggested six-section template and Four Pillars framework
Training navigation and instruction-version selector.
  1. Highlight 1: The Train tab.
  2. Highlight 2: Agent Instructions, Documents and Advanced Settings.
  3. Highlight 3: Instructions Version: which saved version you are looking at.
Instruction text and save controls.
  1. Highlight 1: Current Instructions: edit the text here.
  2. Highlight 2: Save
  3. Highlight 3: Save version keeps this text as a new version.
  1. Skills tab: Enable communication channels (start with Live Chat only), configure the Escalation Skill, and set up any Functions needed
Skills navigation and add control.
  1. Highlight 1: The Skills tab.
  2. Highlight 2: The plus button adds a skill from Skills to add.
Existing skill's edit control.
  1. Highlight 1: The pencil opens a skill's settings (here, Escalation).
Skill setup-status row.
  1. Highlight 1: Each row says whether the skill still needs setting up.

In raia Command:

  1. Create a project for your agent's knowledge
  2. Upload documents (Markdown preferred, with short, clearly headed sections)
Upload with AI drawer: Destination folder and Convert with AI before selecting files. The initial file filter still blocks unsupported types, including images.
  1. Highlight 1: Destination folder
  2. Highlight 2: Convert with AI
  1. Convert to Markdown if needed
  2. Send the converted documents to the agent for training, wait until they show as trained, then ask a test question

See the knowledge worked example for each of these steps with screenshots.

Phase 3: Test (Module 6)

In raia Copilot:

  1. Create a Test Script covering common tasks, missing information and refusal or escalation boundaries. Write the expected answer or action for each case and include realistic variations
  2. Run the test conversations
  3. Check Sources on replies that are wrong
Sources lists the files a reply cited. It is not a list of every piece of text that was searched.
  1. Highlight 1: Click Sources under the reply.
  2. Highlight 2: The Files tab.
  3. Highlight 3: The cited file: northstar-support-guide.md.
  1. Use the Diagnostic Framework as an investigation guide, not root-cause proof: review knowledge and retrieval for unexpected citations, instructions and source use for incorrect answers, and configuration, access, and execution for missing actions
  2. Iterate: fix the source, re-test, confirm the fix
  3. Before a controlled rollout, review results against your acceptance criteria and confirm the agent took no unintended actions

Phase 4: Deploy (Module 7)

  1. Start with Live Chat on your website (copy embed snippet from Launch tab)

Worked example

The examples use fictional data. Follow the same checks with approved, non-sensitive information for your own task.

Open the agent, then Launch. Agent Status shows Public, Private or Inactive; Private means only organization members can chat with this agent. View Embed Code gives the snippet for your site. Select Test Agent to try a sandbox conversation. If no chat appears, keep the agent Private and report the issue; switching to Public broadens access and is not a preview fix. Changing the status changes who can reach the agent, so check with your admin first.

Launch tab for a Private agent.
  1. Highlight 1: Launch tab
  2. Highlight 2: View Embed Code
  3. Highlight 3: Agent Status
  1. For example, plan 2–4 weeks of monitoring in Copilot; adjust based on observed results
  2. Expand to additional surfaces: Browser Extension, Teams/Slack, SaaS integrations
  3. Add SMS, Email, Voice channels one at a time as confidence grows

Phase 5: Integrate (Modules 8–9)

  1. Configure authorized CRM access via webhooks (for example, onThreadEnd → log to HubSpot/Salesforce), then test the workflow
  2. Configure and test Functions with authorized access for live data retrieval (order status, account balance)
  3. Set up n8n ETL pipelines to send updated documents, then confirm each upload, training and a test answer
  4. Expose the agent via MCP for developer tool integration
  5. Build custom UIs via the API if needed

Phase 6: Improve (Continuous)

  1. Monitor conversations daily in Copilot
  2. Provide feedback (especially Suggested Replacements) on bad answers
  3. Review the Scoring Skill using your rubric; below 6 is an example threshold from the demo's custom 1–10 rubric
  4. Update Instructions and Vector Store based on patterns you observe
  5. Save instruction versions so you can compare a tested change and go back if needed

Example First-Week Plan

The dates, document counts, and test counts below are illustrative planning inputs, not requirements. Proceed to production only after actual acceptance checks, permissions, channel tests, and side-effect checks support readiness.

DayActivityProduct
Day 1Answer 5 Planning Questions, draft objectives and guardrails—
Day 2Draft instructions using the suggested template if useful, create agent in CXraia CX
Day 3Format and upload, for example, 10–20 relevant documents to Vector Storeraia Command
Day 4Run, for example, 15 test conversations, fix issues and assess acceptance checksraia Copilot
Day 5Enable Live Chat on a staging/internal page, invite team to testraia Chat
Day 6–7Monitor, gather feedback, refine instructionsraia Copilot
Week 2If readiness checks pass, begin a controlled production rollout; otherwise continue fixing and testingAll

Final Principle: An AI agent is never "finished." Treat agent management as a continuous loop: Monitor → Spot Issues → Fix the Source → Re-test → Repeat. Feedback informs review; it does not automatically retrain the agent or guarantee permanent improvement. Verify changes and keep monitoring.

Practice check

Recommended practice: trace one support question from its source document to the reply. Record whether the answer and any requested action were verified separately; do not mark a real action successful based only on the agent saying it happened.

Previous lesson