LangSmith can now hide trace inputs and outputs from engineers while still showing them latency, cost and run trees
Agent traces are full of customer data, and LangSmith now lets admins decide who sees that part.
What happened
LangChain announced three new access controls in LangSmith on November 17, available on Enterprise plans. The first restricts who can see trace inputs and outputs, the user messages and model responses where sensitive data usually sits, while still showing run trees, latency, token usage and metadata. The second adds user groups, each with its own permissions per workspace, layered on top of existing role-based and attribute-based access control. The third lets admins write trace-level policies against root run metadata, for example giving a support team access only to traces for the customer IDs it handles.
The docs describe the permissions as runs:read, runs:read-inputs and runs:read-outputs, so a user can browse runs in a project that holds PII without seeing what was in them. LangChain's example has clinicians who can view inputs and outputs for annotation, and agent engineers who see only run trees, latency and token usage.
LangChain frames this as an alternative to duplicate traces across workspaces or not sending production traces at all, and cites GDPR, HIPAA, CCPA and customer contracts as the reasons access has to be limited.
My take
Tracing is one of the first things I add to a client's agent, and it creates a second copy of their customer data that is easy to forget. Debugging a slow tool call rarely requires reading the customer's message, so hiding inputs and outputs by default for engineers is a sensible place to start.
The metadata policies only work if you tag every trace at the root with something like a customer or tenant ID, so that tag belongs in the agent code from day one. It is much harder to add after a few months of untagged traces. I'd still redact secrets and payment details before anything leaves the app, because access control limits who sees the data but the trace store still holds it. These controls are Enterprise only, which smaller teams should check before planning around them.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
