Skip to content

Architect track

AI agents & Now Assist

Designing agentic workflows that are safe, measurable and supportable.

2 sections · 13 entries

Agentic workflow anatomy

How AI Agent Studio pieces fit together. Names shift between releases (AI experiences are being rebranded ServiceNow Otto from Australia), so confirm in your release notes.

  1. TriggerA record event, a schedule, or a user in chat starts the work.
  2. Agentic workflowThe goal and plan: which agents collaborate, in what order.
  3. AI agentRole, instructions and the tools it may use.
  4. ToolsFlows, subflows, scripts, catalog items, search / knowledge retrieval.
  5. Human in the loopApproval or review steps before anything irreversible.
  6. Govern & measureGuardrails, logging, usage and outcome metrics, AI governance reviews.

AI design rules

The difference between a demo and something you can run in production.

  • ✦Narrow tools. Give agents small, well-named tools with clear inputs; never a "run any script" tool.
  • ✦Approval gates. Require a human approval before writes that cost money, notify customers or change prod.
  • ✦Least privilege. Agents act as a user with only the roles their tools need.
  • ✦Grounding. Answer from curated knowledge and records; keep knowledge articles current.
  • ✦Evaluate. Keep a test set of real (masked) cases and re-run it after every prompt or tool change.Hidden gem
  • ✦Observability. Log each decision, tool call and outcome so support can explain what happened.
  • ✦Custom skills. Use Now Assist Skill Kit for custom generative skills instead of raw API calls from scripts.Hidden gem

Snippets are provided as-is under the MIT licence, without warranty. Test in a sub-production instance first. ServiceNow is a trademark of ServiceNow, Inc.; this page is independent and not affiliated with or endorsed by ServiceNow.