Trace Screening: Automatically restrict personal or sensitive traces

Last updated: October 27, 2026

Trace Screening automatically finds personal or sensitive agent traces and restricts them so only their author can see the trace contents, regardless of other people's permissions. Once a trace is screened, Span doesn't expose its contents to anyone or anything other than the author, including analysis such as evaluations, coaching, and Harness health. It combines AI-generated labels with screening rules set up for your organization.

Why it matters

By default, trace contents are visible only to the engineer who authored the trace. Some organizations expand that through Agent Trace permissions, so admins or managers can view trace contents too.

But not every agent session is about code. People use agents for performance reviews, board decks, and other sensitive work, and they don't always remember to turn on privacy mode in auspex, which stops capture entirely.

Trace Screening is a safety net for those moments:

For leaders and executives: a session about a performance review or board deck is restricted to its author, even from people who could otherwise view it.
For engineers: added confidence that sensitive activity is identified and restricted automatically.
For admins: a safer way to roll out trace collection across the whole organization.

How it works

Span labels each trace in the background. A set of default labels covers topics such as personal matters, conversations about specific people, and confidential company matters. Your Span Implementation Engineer or Customer Engineer can add custom labels for your organization, described in plain language (for example, "any session about personal finances").
Your screening rules decide which labels restrict a trace.
When a trace matches a rule, it is restricted. Only the author can view its contents.
Span also keeps the contents of a restricted trace out of any analysis that reads trace contents, such as evaluations, coaching, Harness health, and other analysis Span runs on your traces.
New traces stay hidden until screening finishes, so nothing sensitive is visible in the meantime.

What a screened trace looks like

If you're the author, the trace is marked Restricted. You'll see which screening label matched and a short explanation of why. You can still view the full trace.

If you're anyone else, including people with Agent Trace Data or Agent Trace Contents permissions, the trace shows a lock icon. You can see that the trace exists, who authored it, and its model, cost, and duration. Prompts and the transcript are hidden.

Screened traces can't be filtered or searched as a group. This is intentional, so sensitive sessions aren't easy to go looking for. A search across traces also hides the contents of restricted traces.

Removing or restoring a restriction

In the Span app, only the trace's author can change its restriction. If a trace was screened by mistake, the author can remove the restriction:

Open the trace, or find it in the Traces list.
Open the actions menu (the three dots) and select Remove restriction.
Choose a reason and confirm.

To restrict the trace again, the author opens the same menu and selects Mark restricted. Either change takes a few hours to apply, and until then the trace keeps its current state.

Setting up Trace Screening

Trace Screening is off by default, so nothing changes for your organization until you choose to turn it on.

To get started, reach out to your Span account team. Your Span Implementation Engineer or Customer Engineer will create and maintain your screening labels and rules with you. Span manages labels because a broadly worded label can match far more traces than intended, and careful setup protects your data.

Things to know

Metrics still include screened traces. Trace metrics are calculated from trace metadata, such as how many prompts a session had, its cost, and turns per line of code. None of them read trace contents, so they still count screened traces. For example, if your organization ran 200 sessions this week and screening restricted 6, session counts and cost include all 200, while no one but each author can read those 6 sessions, and analysis that reads contents, such as evaluations, skips them.
Screening applies going forward. Rules apply to newly collected traces. Historical traces are not rescreened.
New traces can take a few hours to appear. Because traces stay hidden until screening finishes, there may be a delay before a new trace shows up in Span.
Very short sessions skip screening. A session with at most one prompt, no tool calls, and a duration of 60 seconds or less isn't screened. Use privacy mode for a quick sensitive question.
It's a best-effort safety net. Labels are AI-generated, so occasionally a trace may be restricted that shouldn't be, or a sensitive trace may be missed. If you notice gaps, your Span Implementation Engineer or Customer Engineer can refine label definitions.
The whole trace is labeled. Screening applies to the entire trace, not individual messages.
It works after collection. Screening restricts traces that have already been collected. To keep a session from being collected at all, use privacy mode: while it's on, auspex captures, stores, and sends nothing from that machine. Your organization can also turn on privacy mode for every device through your Span account team.
Engineers can't flag a trace themselves. Users can't mark an unscreened trace as sensitive. Traces can be deleted by people with permission to manage them, and your organization's retention policy also applies.
Label changes take a few hours. Adding, editing, or removing a label takes a few hours to apply.

Need help?

Contact your Span account team or reach out to support@span.app.