Skip to main content
Back to timeline
langchain.comSource publication:

LangSmith adds three access gates to production traces: engineers see run trees, not customer names or payment details

Synopsis

LangSmith has added three access controls to its Enterprise plans for production traces: admins can hide trace inputs and outputs, assign permissions by user group, and set trace-level policies from root run metadata. Previously teams either built custom observability setups with duplicated traces and multiple workspaces, which cost platform teams more to configure and maintain, or kept traces out of observability platforms, which left agent engineers without visibility. The controls layer together so engineers can inspect run trees, latency and cost while sensitive inputs and outputs stay with the people who need them.

Source-provided article image: Protect sensitive data in your traces: new controls in LangSmith
langchain.com

Interpretation

LangSmith Enterprise now offers three controls over who can see sensitive content in production traces: input and output access control, permissions managed by user group, and trace-level access policies driven by root run metadata. Teams previously had two options: a custom observability setup with duplicate traces and multiple workspaces, which left platform teams more to configure and maintain, or not sending traces to observability platforms at all, which removed agent engineers' visibility and slowed product velocity. The three controls are available in LangSmith Enterprise plans, and the stated pattern is to send traces with full data fidelity and let these controls determine what each team can view.

The design separates authorization for sensitive content from the structural information needed for debugging: run trees, latency, cost, token usage and metadata remain available to engineers, while inputs and outputs are restricted separately. Inputs and outputs are where sensitive user data most often resides, whereas run trees, performance stats and metadata are critical for agent improvement and far less sensitive. The stated example puts agent engineers in a group that can view only run trees, latency and token usage, and clinicians in a group that can view inputs and outputs for accurate trace annotation.

The finest layer is trace-level attribute-based policy: admins use conditions on root run metadata to decide which traces a user can see, such as customer IDs limiting support agents to the customers they support or user IDs limiting developers to their own coding agent traces. Group permissions are defined per workspace and layer on top of the organization's existing role-based and attribute-based access controls, so existing restrictions such as regional access policies are preserved. The stated combined scenario has a support team viewing inputs and outputs only for its customers while agent engineers view run trees and performance data across all traces with inputs and outputs hidden; when the support team takes on another customer, only the metadata policy changes.

Perspective

The controls are currently available in LangSmith Enterprise plans, and the stated setting is sending complete traces to LangSmith and then scoping visibility by team; the text does not say whether other plans offer them, nor does it give deployment scale or outcome data.

The controls are presented as useful individually and designed to be layered, but no results from actual deployments or audits are reported; metadata-based policies require the relevant attribute to be captured in root run metadata.