Claire_Conant
ServiceNow Employee
Options
- Post History
- Subscribe to RSS Feed
- Mark as New
- Mark as Read
- Bookmark
- Subscribe
- Printer Friendly Page
- Report Inappropriate Content
on 06-12-2026 01:18 PM - edited 3 weeks ago
Setting up AI agents goes smoothly until it doesn't. You’ve activated the Now Assist AI agents plugin, opened AI Agent Studio, and you’re ready to start building, then something goes wrong. The product docs cover the setup steps, but there are several configuration behaviors that can stall first-time deployments. Most of them don’t produce obvious error messages, which makes them harder to spot.
The situations below are the ones that come up most often.
Does this apply to you?
This article covers configuration behaviors specific to initial AI agents setup, beginning with the Zurich release. If your agents are already running in production and you're seeing runtime failures, tool execution errors, or response quality issues, the general troubleshooting guide is a better starting point.
|
What you're seeing
|
Where to start
|
|
Prebuilt agent hangs at "Task Start" with a 500 error in the console
|
|
|
Agents stop responding, skills go missing, or AI Agent Studio throws errors after a patch
|
|
|
LLM provider setting quietly reverts to a previous value
|
|
|
"Security constraints prevent access" error in AI Agent Studio or Skill Kit
|
|
|
Provider not listed during configuration, or agent fails silently after saving
|
|
|
Custom agent is configured and available but never responds to prompts
|
1. Prebuilt agents get stuck at “Task Start”
This one tends to catch people early. You open AI Agent Studio, try to test a prebuilt agent, and it just hangs at the “Task Start” step. You might see a generic 500 Internal Server Error in the browser console, but it doesn’t point you anywhere useful. Repairing or reinstalling Now Assist for ITSM doesn’t help either.
What you can do
An inactive assistant is usually the cause. The setup docs don't call this out as a prerequisite, so there's no reason you'd think to check it. To verify, go to Conversational Interfaces > Assistant Designer > Assistants. Review the setup and activate the assistant associated with your configuration, then retest your agent.
2. Agents stop responding or throw errors
After a patch or upgrade, AI agents can start behaving unpredictably: agents don’t respond, skills aren’t discovered, or the AI Agent Studio throws errors. This can happen if you update individual child plugins instead of starting from the parent.
What you can do
AI agents depend on a chain of plugins and updating them out of order can leave the dependency chain in an inconsistent state. The safer approach is to always update the parent plugin. For example, updating “Now Assist for ITSM” cascades updates to all its dependent child plugins. It’s worth checking for updates after every patch.
3. LLM provider settings can silently revert after an upgrade
You configure your preferred LLM provider through Now Assist Admin, and everything looks right. But within a few hours, the setting quietly reverts to its pre-upgrade value and you might not notice until an agent starts failing or producing unexpected results. This can happen if you upgrade your instance to a new release but don’t also update the Generative AI Controller and Now Assist Admin apps to the latest compatible versions. The older app versions silently reset the provider configuration.
What you can do
Update both store apps to the latest versions compatible with your release before configuring LLM provider settings. If you're on Zurich Patch 6 or earlier, upgrade to Patch 7 or later to resolve this behavior permanently. This isn't the only thing that can go sideways after an upgrade; if AI Agent Studio itself stops loading, the next section covers that.
4. AI Agent Studio stops working after a plugin update
Unlike the silent provider revert above, this one is harder to miss. After updating the Now Assist AI Agents store app, AI Agent Studio or Now Assist Skill Kit can return a "Security constraints prevent access to requested page" error. This happens when the Now Assist Admin Console store app hasn't been updated to a compatible version.
What you can do
To resolve this, update the Now Assist Admin Console store app to version 6.6.4 or later.
5. A newly activated instance leaves agents and flows without access by default
A version mismatch isn't the only reason you can end up without access to AI agents. If the instance you're working on was created fresh or fully reset to base system defaults on Brazil—rather than upgraded or patched—the cause is ACLs, not app versions. Five AI ACL types, gen_ai_agent, gen_ai_workflow, gen_ai_skill, flow, and flow_action, start from a deny-by-default configuration on new and freshly reset instances. This is a change from permissive wildcard access these types used to inherit. Now each of these wildcard ACLs denies access until another specific ACL entry grants access to a given AI agent. An instance that's only been patched or upgraded in place keeps its prior configuration and isn't affected.
What you can do
Open the guided setup for the component you're configuring: "Define security controls for an AI agent" for agents, or "Define security controls for an agentic workflow" for agentic workflows. Each setup offers three options: any authenticated user, users with specified roles (the default), or public access.
Flows and flow actions work differently. There's no guided setup; you create an Allow If access control using the invoke_from_ai operation. The deny-by-default only applies when a flow is invoked by AI; flows triggered any other way aren't affected.
Note: In Brazil Patch 0, deny-by-default for flow and flow_action was temporarily reverted to permissive wildcard access. It's restored as of Patch 1. If you're on Patch 0 through the Early Availability program, you won't see this change yet.
6. Your model provider may not be available for every agent or skill
Not all four model providers (Now LLM Service, Azure OpenAI, Google Gemini, Anthropic Claude on AWS) work with every skill and agent type. During configuration, you may have selected a provider only to find it's not in the list, or the agent fails silently after you save. Certain skills and agentic workflows have provider restrictions that aren’t always visible during setup.
What you can do
Before configuring a provider, check the support matrix to confirm it’s supported for your specific agent or skill type, release, and market.
7. Custom agent never responds even though it's configured and available
If your custom AI agent doesn't respond to user prompts in the Now Assist Panel or Virtual Agent even though it's configured and available, the issue might be the agent's name and description. AI agent discovery works by matching what users type against what you've written in these fields. Generic names like "License Processing" or one-word descriptions don't give the system enough signal to route correctly.
What you can do
The fix is straightforward: write names and descriptions that clearly describe the task and context. For example, "Process software license requests for employees" gives the discovery mechanism much more to work with than "License Processing."
Where to go from here
You still have a few more options if your AI agents are working as you expect them to. The general troubleshooting guide walks through diagnostic steps to pinpoint where things are failing.
For questions about runtime behavior, testing in the AI Agent Studio playground, or channel configuration, the AI agents user guide covers those areas.
The good news is that none of these are hard to handle once you know to look for them. Some are quick fixes you'll make once, like activating an assistant. Others are worth folding into your setup routine, like updating from the parent plugin or keeping your store apps in sync after an upgrade. Catching them early means you'll spend less time troubleshooting and more time building agents that solve real problems.
