Claude Code regularly sends the transcript of each background agent to its underlying model, according to a gist published online. The claim drew nearly 400 points and more than 130 comments, making it one of the most debated items of the day.
The detail matters because background agents can run for extended periods and accumulate substantial context. If each transcript is transmitted, users may be sending far more information to the model than they assume when they launch an automated task. The gist presented the behavior as a notable and possibly surprising aspect of how the tool operates.
Commenters split on whether this was a problem. Some argued that sending transcripts is inherent to how such agents function, since the model needs context to act. Others said the issue is not the transmission itself but how clearly it is disclosed to users who may be handling sensitive code or data.
The discussion fits a broader pattern of scrutiny around AI coding tools, which increasingly operate with limited direct oversight. As agents take on longer tasks, questions about what they retain, transmit, and expose are likely to keep surfacing. Anthropic has not publicly responded to the specific claims in the gist.
Natasha Rust