Claude Code burns ~15–20% of your tokens on a 3-word status line

Claude Code's background summarization feature is disabled by default, reducing your token costs.

A user-developed fix, detailed on GitHub by Flamerix AI, addresses a significant token consumption issue in Claude Code. The problem lies in a background process where Claude Code sends an agent's entire transcript to its own model every ~30 seconds to generate a three-word status update. This 'agent_summary' function, which defaults to being on since August 2026, consumes approximately 15-20% of a user's tokens and can even cost more than the main conversation itself, according to OpenTelemetry data. The fix involves a Bun preload script named 'no-agent-summary.js' that intercepts and disables these specific summary timers without altering the core Claude Code binary. The script is configured via environment variables, such as BUN_OPTIONS, and requires restarting the Claude Code daemon to take effect across all sessions.

This discovery reveals a substantial inefficiency in how current AI coding assistants manage internal processes and resource allocation. The 'agent_summary' function, designed for user convenience by providing status updates, inadvertently becomes a major cost driver, highlighting a potential disconnect between developer-intended features and their real-world economic impact. The fact that this is a default setting and difficult to disable underscores a broader challenge in the industry: managing the operational costs of AI agents that operate continuously in the background.

The workaround, while effective, points to a gap in the extensibility and configurability of these tools. Users are forced to resort to preloading scripts to modify core behavior, suggesting a need for more granular control and transparency in AI agent settings. Future developments will likely focus on either more efficient summarization techniques or providing users with explicit controls to manage these background processes, potentially impacting the pricing models and user experience of AI coding assistants.