The Pulse
LangSmith adds trace access controls for Enterprise customers
LangSmith Enterprise customers can restrict access to trace inputs and outputs, set group permissions and control access using metadata. The controls govern who can view stored traces rather than requiring redacted or duplicate copies.

LangSmith Enterprise customers can now restrict who sees the inputs and outputs recorded in AI agent traces while allowing other users to inspect operational data. LangChain says the controls let engineers debug agents without necessarily seeing sensitive messages or model responses.
The new controls cover input and output visibility, permissions for user groups, and access policies based on trace metadata. They are available in LangSmith Enterprise plans.
Separating content from debugging data
Traces record how an agent behaved, including inputs and outputs such as user messages and model responses. Those fields can contain customer information, including names, health information or payment details.
The input and output controls let administrators limit access to those portions of a trace while leaving other information visible. For example, an engineer could inspect run trees and performance statistics such as latency and cost without seeing the underlying messages.
LangChain says engineers can use that operational information to investigate issues such as repetitive or slow tool calls. Users who need the complete picture can be given access to the inputs and outputs.
Permissions for groups and traces
Administrators can organize users into groups and set permissions for each workspace. Those permissions layer on top of existing role-based and attribute-based access controls across the organization, while preserving existing restrictions such as different policies for employees in different regions.
The company gives the example of a group of clinicians allowed to view inputs and outputs for trace annotation, alongside an engineering group limited to run trees, latency and token usage. The groups can therefore be assigned access according to their roles.
Admins can also define trace-level attribute-based policies using metadata on the root run. Each policy checks a trace’s metadata against conditions set by the organization.
For instance, a policy could use customer IDs to give support agents access only to traces for customers they support. It could also use a user ID to limit developers to their own coding agent traces, provided the relevant attribute is captured in root run metadata.
Stored traces and masking
The controls can be combined. A support team could see inputs and outputs for its customers, while engineers could see run trees and performance data across traces with those inputs and outputs hidden. A policy for the support team could then be updated for another customer without changing the team’s input and output permissions or the engineers’ access.
LangChain says complete traces can be sent to and stored in LangSmith without requiring redacted or duplicate copies; the controls govern what each team can view. That approach differs from removing information before it is sent.
Separate LangChain documentation describes ways to protect data before it reaches the backend. Those options include hiding inputs and outputs, transforming metadata, and using rule-based masking or anonymization to redact selected information. The documentation notes that anonymization may affect performance with complex regular expressions or large payloads.
Sources
Companies in this story
Related stories
- LangChain puts agent records at the center of LangSmith
- Microsoft introduces enterprise policy controls for AI agents
- Firecrawl kit gives Docker sandbox agents controlled web access
- Microsoft and partners start enterprise AI transparency initiative
- Appian expands agent control tools, with more features planned