Collibra now reads agent quality signals from four AI platforms. A shared dashboard is not a shared metric

Suppose an agent on Amazon Bedrock passes 92% of its quality checks this week and an agent on Databricks passes 88%. Which one is more reliable?

There is no way to know from those two numbers, and that is the useful question to bring to Collibra’s latest release.

Collibra Platform 2026.10 reached commercial pre-production environments on October 11, with production scheduled for November 8. It extends what Collibra calls operational trust: signals about what deployed agents actually do, collected from the platform each agent runs on and shown alongside governance information such as assessments, lifecycle status and documentation completeness.

What it collects

Two kinds of signal. Monitor pass rates: Collibra records the pass or fail result of each monitoring check performed against an agent and computes daily and seven-day pass rates. Token consumption: prompt and completion token counts. These appear on a Quality tab for each agent and each agent version, and on an Agents dashboard.

The capability is available for agents integrated through Databricks, AWS Bedrock, Gemini Enterprise Agent Platform and Azure AI Foundry. It must be switched on, and it depends on the underlying platform having monitoring data, such as interaction traces in Databricks or CloudWatch invocation metrics in Bedrock. Without that data, nothing appears.

For enterprises whose agents live on several clouds, this is a real step. Most governance tools have been inventories; this one reads behavior. We described the market for control planes above the vendors earlier this month, and operational evidence is what separates a control plane from a catalog.

The sentence that matters

Collibra’s documentation includes a caveat that deserves to be on the dashboard itself: “What is collected, and how a pass or fail result is determined, differs by AI integration.” The number of monitors linked to an agent and how often current-day figures refresh also vary by platform.

So the 92% and the 88% may measure different things: different checks, different judges, different thresholds, different sampling. Placed side by side on one screen, they look comparable. They are not, unless someone has made them so.

This is not a flaw specific to Collibra. Any layer that aggregates signals from platforms it does not control inherits their definitions. As we wrote about agent observability beyond uptime, the value of a quality metric is in its definition, and an aggregate is only as consistent as its least consistent source.

What to ask before trusting the comparison

Start with coverage. The Agents dashboard includes an evaluation coverage gap: agents with no LLM judge assigned, which, the documentation notes, contribute no operational trust data. That is the most honest number on the screen. An agent with no monitors does not have a 100% pass rate; it has no evidence.

Standardize a few checks. Comparability comes from running the same evaluation on agents across platforms, for example a shared set of test cases for groundedness or policy compliance, not from collecting each platform’s native metric.

Keep observed and computed apart. Collibra’s AI Trust Score blends signals. Ask to see the operational components separately from the governance ones, so that good documentation cannot hide poor production behavior.

Use token data for owners, not only for dashboards. Token consumption per agent version is the raw material for the budgets every production agent needs, the record we put in every agent’s file.

The same release adds, in preview, the option to connect an organization’s own Gemini Enterprise credentials as the model provider for Collibra Maestro, Collibra’s AI assistant. It is a separate feature, but it points in the same direction: governance tools are starting to operate inside the customer’s AI estate rather than beside it. That makes their definitions matter more, not less.