A gist published online alleges that Claude Code, Anthropic's command line coding tool, regularly sends the complete transcript of every background agent it runs to its model. The claim, posted by a user identified as flamerix-ai, has become one of the most discussed items in developer circles, drawing 451 points and 183 comments on Hacker News.
The core concern is scope. Background agents are designed to work autonomously on tasks, and if their full transcripts are shipped to the model on a regular cadence, developers want to know exactly what content is included, how often it happens, and whether anything sensitive from a local codebase or conversation history travels with it. The gist does not appear to allege malicious behavior, but it does raise transparency questions about how agent context flows through the system.
Anthropic has not issued a public response to the gist at the time of writing. The discussion thread reflects a broader unease in the developer community about autonomous tooling, where agents may read files, run commands, and accumulate long histories that users never explicitly review. Many commenters asked for clearer documentation on what gets transmitted and how it is retained.
The story lands as AI coding assistants become deeply embedded in professional workflows. Claude Code competes with similar tools from other vendors, and trust around data handling has become a differentiator. Developers say they want opt-in controls and visibility into agent telemetry rather than having to reverse engineer behavior from network traffic or logs.