All work

Case study · AI agent observability

Observability for AI agents: streaming Zuora activity to your security stack

Customers now change Zuora through MCP clients, whether AI agents or people working in tools like Claude. I product-managed the connector that captures those changes and streams them to the monitoring stack security teams already run.

Role
Product manager for the initiative
Company
Zuora
Domain
AI agent observability, audit event streaming
Destinations
Splunk, Datadog, Sumo Logic, or any tool via API
3
native connectors, plus any tool via API
5 min
fastest push interval, near real-time
0
lines of code to configure or maintain

The problem: changes arrive through new doors, and security can't see them

Once customers connect MCP clients to Zuora's MCP servers, changes to the system can come from an AI agent or from a person using a tool like Claude, not only from the product UI. Security teams already centralize detection in their monitoring stack, but audit data for the platform lived behind manual exports and one-off integrations. The gap between a change happening and a security team knowing about it was measured in exports, not minutes.

Before

  • Manual exports to track logins, settings and object changes
  • Audit data scattered across platform and billing tenants
  • Slow, brittle custom integrations to build and maintain
  • Delayed detection of security and compliance events

With the connector

  • Events pushed automatically, with no manual export step
  • One destination for platform and billing tenant activity
  • Configured once in the admin console, no code to maintain
  • Near real-time delivery, as often as every 5 minutes

How it works

Three steps, all from the OneID admin console, and then it runs on its own. It streams authentication and access events, settings changes and record-level changes across platform and billing tenants. That includes the changes made through MCP clients, so agent-driven and human-driven activity land in the same stream.

01

Configure a connection

Pick the monitoring tool, enter the endpoint URL and credential, then run Test Connection before saving. A connection that cannot reach its destination never gets saved.

02

Create a schedule

Choose the connection, the event types, the tenants to include and a push interval anywhere from every 5 minutes to once every 24 hours.

03

Zuora streams the events

Each interval, matching events are pushed to the customer's tool. Every schedule can be monitored, paused or resumed, with last and next push times visible.

Product decisions

Send security teams signal, not noise

Schedules filter by event type and by tenant, so a security team receives only the activity it actually needs to watch, rather than the full audit firehose.

Meet security teams in the tools they already run

The destination is the customer's own monitoring stack. Nobody has to learn a new console to see what agents and MCP users are changing, and detection and alerting stay in the workflows security teams already trust.

Make failure visible before it happens

Test Connection validates the endpoint and credential up front, and the schedule list shows last and next push times, so a silent gap in monitoring is hard to create and easy to spot.

State the boundaries plainly

A schedule streams only events generated inside its interval window. Events from before it was activated, or while it was paused, are not pushed, and older entries come from the Audit Trail APIs. Predictable semantics beat hidden buffering.

Connect to the tools you already use

Native connectors stream directly to Splunk, Datadog and Sumo Logic. Because events are also available through the Audit Trail APIs, customers can feed any other tool in their stack, such as CrowdStrike, Microsoft Sentinel or Elastic, without waiting for a dedicated connector. Additional native connectors are planned.

  • Splunk
  • Datadog
  • Sumo Logic
  • CrowdStrike
  • Microsoft Sentinel
  • Elastic
  • Any tool via API

Where it fits

Authorization decides what an MCP client may do. Observability shows what it did. Together with OneID's MCP authorization, this closes the loop: identity and access controls on one side, and a live record of every change in the customer's own security stack on the other.

More event types are planned, so the connector grows with the tools security teams adopt.Read the OneID case study →

Working on agent security or observability?

I'm always glad to compare notes on securing and monitoring AI agents in enterprise products.

Get in touch