A telemetry and sensitive information filtering scheme for coding agents

On November 7, 2026, at 10:44 a.m., DEV Community's AI channel published an article titled "Telemetry and sensitive-information filtering for coding agents". The author describes evaluating MLflow, Langfuse and LangSmith for use with his own coding agents. All three tools readily record spans, and two of them even evaluate outputs with an LLM judge, so the shortfall he found was not in tracing itself but in workload fit. A coding session, he argues, is not a single LLM call: it is a tree of tool calls, file edits, retries and nested agents that can run for hours. The moment a team decides to store transcripts, redaction stops being a footnote and becomes the storage design. On that basis, the article proposes a tracing scheme that automatically filters sensitive source code and environment information before anything is written to storage. In the cited passage, the author points to the telemetry stream containing source code and .env content as evidence of the risk, but the specific cases behind that claim, and the details of how the filtering scheme is implemented, are not covered in the reporting available here.