- Write clear action descriptions, because they decide when the agent calls each tool
- Keep knowledge sources focused and up to date
- Start a New test session before each test run so earlier turns do not skew results
Knowledge
- Split large documents into focused sources (returns policy, shipping, billing) so retrieval finds the right passage
- Use descriptive titles on Text sources to make debugging easier
- Re-crawl or re-upload when content changes, since stale chunks cause outdated answers
- Remove deprecated content instead of leaving conflicting sources in place
Actions
- Treat the description as “when to call”. Write it like an instruction to a colleague: “Search records by reference when the user asks about order or ticket status.”
- Minimize required parameters. The agent asks follow-up questions for missing fields, but fewer fields means faster resolution.
- Test with the action test panel on the action detail page before enabling it in production.
- Never store secrets in dashboard-visible URLs. Use saved tokens or a server-side proxy instead.
Instructions & branding
- For Customer Support agents, instructions are generated from your sources. To change behavior, add or improve documents rather than editing instructions directly.
- Set a welcome message and 3 to 5 suggested prompts that mirror real user intents.
- Match the tone to your brand, whether that is formal support or a casual assistant.
Before go-live
1
Run guardrails
Configure topics, PII filters, rate limits, and frustration escalation under Agent settings > Security.
2
Test edge cases
Start a New test session and try out-of-scope questions, multi-turn slot-filling, and escalation phrases.
3
Configure handoff
Add an Escalate to human action. If you use WhatsApp, configure handoff under Actions > Integrations.
4
Deploy to staging
Use the demo links on the Deploy > Chat widget > Embed tab (website widget or inline panel), or a sandbox WhatsApp number, before sending production traffic. Confirm your test messages appear in Chat logs.
Guardrails
Safety controls and fallbacks.
Use-case recipes
Starter patterns for common deployments.