ServiceNow AI agents

Discovery (read-only). Every agent in AI Agent Studio with who can start it (public, any signed-in user, specific roles, Virtual Agent chat, email or record triggers), who it runs as and that user's roles (admin, security_admin and other elevated roles are flagged), each tool and what it invokes (script, record operation, subflow, flow action, catalog item, web search, knowledge search, MCP), and whether a person confirms the tool first (Supervised) or it runs on its own (Autonomous).

Risk reasons include an agent people outside IT can start that can change records, a run-as user with admin roles, script tools (more when the script evaluates text as code or switches users), tools that change records without a person confirming, and tools that change records without calling the Intertrace guard.

Runtime guard. ServiceNow runs the agent's model inside the instance and has no outside pre-tool hook, so each tool that changes records calls a Script Include first. It asks the enforcer (POST /v1/actions/authorize) and the tool goes on only on allow.

1. 1

   Turn on the client credentials grant

   System Properties: create glide.oauth.inbound.client.credential.grant_type.enabled = true (global scope).
2. 2

   Create the OAuth application

   System OAuth → Application Registry → New → OAuth API endpoint for external clients. Set the grant type to Client Credentials and pick an OAuth Application User: a dedicated, non-interactive integration user. On newer releases this lives under Machine Identity Console → Inbound Integrations.
3. 3

   Give the user read roles

   sn_aia_viewer reads every AI Agent Studio table. For who-can-start and run-as roles, the user also needs read on sys_security_acl_role, sys_user, sys_user_has_role, sys_user_role and sys_agent_access_role_configuration; reading ACL roles usually needs security_admin or a read ACL you add. Without them the source shows as limited and names each table. Intertrace never writes.
4. 4

   Connect in Intertrace

   Discovery → ServiceNow AI agents → Connect: the instance URL (https://<instance>.service-now.com only), client ID and secret. Intertrace gets a token and reads one agent row before saving; the secret is encrypted at rest.

Tables read: sn_aia_agent, sn_aia_agent_config, sn_aia_agent_tool_m2m, sn_aia_tool (the script is read to classify the tool and is not stored), sn_aia_team_member, sn_aia_usecase, sn_aia_trigger_agent_usecase_m2m, sn_aia_trigger_configuration, the ACL and role tables above, and the property sn_aia.enable_usecase_tool_execution_mode_override.

1. 1

   Mint the runtime key

   Import the agent from Discovery, open its agent page and create a runtime key (itr_rt_…). The agent page's id is what the guard sends as agent.id.
2. 2

   Store it in a Connection & Credential alias

   An HTTP connection with Connection URL https://intertrace.fly.dev and an API Key credential holding the runtime key (a Basic Auth credential's password also works). Allow outbound intertrace.fly.dev:443 if the instance restricts hosts.
3. 3

   Add the Script Include

   Create Script Include IntertraceGuard from runtime-plugins/servicenow-intertrace/IntertraceGuard.js and set the property intertrace.guard.connection_alias to the alias's sys_id. Map ServiceNow agents to Intertrace agents with intertrace.guard.agent_map or pass agentId in each call.

Script tool:

(function (inputs) {
  var verdict = new IntertraceGuard().check({
    agentId: 'INTERTRACE_AGENT_ID',
    tool: 'reassign_incident',
    operation: 'write',
    arguments: { number: inputs.number, assignment_group: inputs.assignment_group },
    purpose: inputs.reason
  });
  if (!verdict.allowed) return verdict.message;
  // ... the tool's own work
})(inputs);

Subflow or flow action tool: a Flow Designer action Intertrace guard (inputs agent_id, tool, operation, arguments_json, purpose; outputs allowed, message) with this Script step, as the first step, and an If that ends the subflow with message when allowed is false:

(function execute(inputs, outputs) {
  var verdict = new IntertraceGuard().check({
    agentId: inputs.agent_id,
    tool: inputs.tool,
    operation: inputs.operation || 'write',
    arguments: inputs.arguments_json,
    purpose: inputs.purpose
  });
  outputs.allowed = verdict.allowed === true;
  outputs.message = verdict.message;
})(inputs, outputs);

A record-operation (CRUD) tool runs a script ServiceNow generates: replace it with a script tool that calls the guard first. Keep tools that change records on Supervised as well; ServiceNow's confirmation and Intertrace's approval are separate.

Discovery reads a script tool that calls IntertraceGuard as guarded. It can't see inside flows; the agent shows as protected once the guard's first check reaches Intertrace.

sn_aia_execution_plan (one per request), sn_aia_execution_task (its steps), sn_aia_tools_execution (tools run, kept 13 months) and sn_aia_message hold what the agent did inside ServiceNow. Intertrace records each check the guard makes on the agent's Activity and in Runtime.