LangSmith Adds Controls for Sensitive Traces

LangSmith Adds Controls for Sensitive Traces
Wednesday, November 18, 2026
  • •LangChain added three LangSmith controls for access to sensitive production traces.
  • •Administrators can restrict trace inputs and outputs, organize permissions by user group, and apply metadata-based policies.
  • •The controls are available on LangSmith Enterprise plans and can be combined while preserving complete traces.

LangChain announced three access controls for sensitive production traces in LangSmith on November 17, 2026. The controls let administrators restrict who can view trace inputs and outputs, manage permissions by user group, and set trace-level policies using metadata. They are available on LangSmith Enterprise plans.

Production traces can contain customer names, health information, and payment details, while engineers use traces to debug agents and improve performance. LangChain says regulatory requirements such as GDPR, HIPAA, and CCPA, along with customer contracts that may require ISO 27001 compliance or SOC 2 audits, can require limits on data access. Teams had used custom observability setups with duplicate traces and multiple workspaces, or avoided sending traces to observability platforms, which reduced engineers’ visibility.

Input and output controls let engineers inspect run trees, latency, cost, metadata, and other performance information while sensitive inputs and outputs remain hidden. Administrators can grant full access to people who need it; for example, access to a customer’s inputs and outputs can be limited to the people or teams handling that customer’s issue.

Administrators can organize users into groups with permissions set for each workspace. These permissions layer on top of the organization’s existing role-based access controls (RBAC) and attribute-based access controls (ABAC). LangChain gives clinicians who need to annotate traces as an example of a group that can view inputs and outputs, while agent engineers may see run trees, latency, and token usage. Existing regional restrictions can remain in place.

For more specific access, administrators can create trace-level ABAC policies based on root run metadata. Examples include allowing support staff to see traces only for their assigned customers, using customer IDs, or limiting developers to their own coding agent traces, using user IDs. Any attribute captured in root run metadata can be used in these policies.

The controls can be combined: LangSmith stores complete traces with metadata such as customer or user ID, without requiring redacted or duplicate copies. Support teams can see inputs and outputs only for customers they support, while engineers can view run trees and performance data across traces with inputs and outputs hidden. Administrators can change a support team’s metadata-based policy when it takes on another customer without changing its other permissions. Existing customers can consult documentation for the controls; prospective customers can contact LangChain.

Subscribe to the AIB Newsletter

The fastest, most reliable AI news, straight to your inbox