Setup
Open Zia Agent Studio from the Zoho One app launcher (requires Zia Enterprise scope, which is included with Zoho One Enterprise). Choose: start from a marketplace template or blank. For learning, start with a template — the Sales Development Agent or Ticket Triage Agent are good first picks because they walk through every configuration surface (data scope, tools, prompts, escalation, notifications) without the blank-slate problem of designing those decisions yourself.
Define the Agent
Name (use a verb-noun convention like “Triage Inbound Tickets” so its purpose is clear in lists), scope (which apps and modules), data access (which views or filters within those modules), target users (who can invoke it), and trigger conditions (chat-invoked, workflow-invoked, or scheduled). An agent needs to know what it does and where it operates. Set scope narrow — one clear job. Expand later. A narrow-scope agent is significantly easier to debug than one trying to handle a wide range of inputs.
Tool Connections
Connect CRM, Desk, Books, People, or external APIs through Zoho Connections. Each tool the agent can call is declared explicitly — there is no “agent has access to everything by default” mode. Permissions flow through Zia; the agent acts as a service account with policy boundaries. Tool granularity matters: enable Desk_Tickets_Read and Desk_Tickets_UpdatePriority separately, not a wildcard Desk_Tickets_Write. The principle of least privilege applies — start with read tools only, add write tools after pilot validation.
Prompt Configuration
Prompts have three layers: System (defines the agent’s persona, scope, and behavior rules — “You are a ticket triage assistant. You categorize tickets by product line and assign priority. You never close tickets.”), Context (data automatically injected per invocation — current ticket fields, related Account, customer ARR), and User (the actual question or task). Edit the System prompt to embed your domain language and explicit constraints. Test prompt changes in Studio’s preview before deploying.
System prompt example for a ticket triage agent:
You are a support ticket triage assistant for [Company].
Your job: classify incoming tickets by product line (Product A, B, C),
set priority (Low, Medium, High, Urgent) based on the rules below,
and assign to the correct queue.
Priority rules:
- Urgent: customer ARR > $50k AND keyword "down" or "outage" in subject
- High: customer ARR > $50k OR keyword "blocked" in subject
- Medium: default for paying customers
- Low: free-tier customers without urgent keywords
Never close tickets. Never reply directly to customers.
If you cannot classify with confidence, escalate to a human reviewer.Test and Deploy
Test in Studio’s preview with sample inputs covering happy path, edge cases, and adversarial inputs (deliberately ambiguous tickets to verify escalation behavior). Verify outputs, check escalation logic, confirm tool calls are within scope. Deploy to a pilot user set of 3-5 people who can review every agent output for the first week. Measure actual outcomes (correct classifications, appropriate escalations, time saved) before broad rollout. Marketplace templates reduce time-to-first-agent to under an hour; a custom agent built from blank typically takes 2-4 hours of iteration before it’s pilot-ready.
Monitoring After Launch
Set credit budget caps per agent. Configure failure alerts (agent timeouts, tool errors, escalation-rate spikes) into a Cliq channel. Review the Agent Activity Log weekly for the first month — patterns of misclassification, repeated escalations on similar inputs, or tool-call loops surface here. Iterate on the prompt based on actual misses, not anticipated ones.
Common Failure Modes
Wide initial scope leading to inconsistent outputs. Missing escalation rules — the agent attempts to handle inputs it shouldn’t. Over-trusting the LLM on edge cases — always have explicit fallback to a human queue. Stale prompts that don’t reflect process changes — quarterly prompt review with the process owner.
What to do this week
Install one marketplace template agent in a sandbox, run 20 representative cases through it, review every output, and document the gaps before configuring for production.
