Claude Code’s Hidden Token Drain

Claude Code’s background agents periodically trigger model requests to generate brief progress labels for the status bar. This behavior, first identified by Flamerix AI AI and corroborated by Better Stack, represents a hidden drain on token usage.

Roughly every 30 seconds, a summary request can include the agent’s full conversation rather than only its latest activity. Better Stack’s telemetry showed these agent_summary calls transmitted up to 40,000 input tokens for a mere 130 tokens of output, such as "running 7th 20-second sleep."

This mechanism impacts users differently depending on their Claude access. For subscription users on tiers like Claude Pro/Max, these frequent, context-heavy requests consume valuable rate-limit capacity, potentially leading to premature throttling.

For API customers, the token consumption translates directly into higher billing. Better Stack’s tests revealed agent summaries accounted for 12% of total session cost in a synthetic benchmark and 8% in a realistic build of a TanStack site with three background agents. These summaries often run on expensive models like Claude Claude Opus, not cheaper alternatives like Claude Claude Haiku.

Better Stack’s OpenTelemetry test confirmed the agent summary behavior. During a background-agent session, five agent_summary requests appeared, spaced about 30 seconds apart. This occurred while a single agent was tasked with eight 20-second sleep cycles.

Each of these calls consumed approximately 40,000 input tokens, generating a mere 130 output tokens for labels like “running 7th 20-second sleep.” This striking input-output imbalance highlights the token inefficiency.

In this synthetic benchmark, agent summaries accounted for 12% of the total session cost. A more realistic test—building a TanStack site with three background agents—showed about 8% of usage dedicated to these summaries, with every summary call running on Claude Claude Opus rather than a cheaper alternative like Claude Claude Haiku.

Flamerix AI AI’s independent telemetry analysis reported an even higher share, with agent_summary requests consuming 16.4% of total spend. Conversely, Anthropic’s internal fleet-wide telemetry suggests a lower 2% of overall API spend and 6% for terminal-only usage, indicating significant variability based on usage patterns.

Overhead from these status updates is highly variable, depending on your specific usage patterns. Factors like the number of active background agents, session length, and the cumulative size of transcripts all contribute. A brief, single-agent task might see minimal impact, but extended multi-agent sessions can quickly escalate costs.

One significant concern is the model choice for these summary requests. Better Stack's telemetry observed agent summary requests consistently using Claude Claude Opus, even when a more cost-effective model like Claude Claude Haiku would suffice for generating a few words of status. This means premium tokens are being spent on trivial tasks.

While Anthropic's internal telemetry reported a modest 2% overall cost and 6% for terminal usage, community measurements often show higher figures, sometimes reaching 8% to 26%. This discrepancy highlights how individual workflows can dramatically alter the impact. Prompt caching, a technique where identical prompts are not re-processed, might also influence reported costs, though its effect on dynamic transcript summaries is less clear. For more details on the tool itself, consult the Claude Code Overview & Documentation.

Two primary workarounds emerged from the community, each with trade-offs. Setting claude code fork subagents to zero eliminates observed summary calls by disabling fork subagents entirely, effectively stopping the token drain. The catch: agents lose the ability to inherit the parent conversation, limiting their utility in complex tasks.

A more technical solution, reported by Flamerix AI AI, uses a Bun preload script to intercept and block the timer responsible for triggering summary requests. While effective, this approach is brittle, relying on internal Claude Code implementation details that could change with updates. It also carries security implications, as the script hooks into every Claude Code process.

Practical advice: inspect and pin any third-party script to a specific version rather than blindly running a remote install command. Monitor your usage via Claude Code’s OpenTelemetry documentation to understand the actual impact on your token consumption. Ultimately, the ideal solution is a native opt-out in Claude Code or a cheaper, dedicated summary model like Claude Claude Haiku for these low-value, high-frequency requests.

Why do Claude Code background-agent summaries use tokens?

Claude Code periodically sends an agent’s conversation context to a model to produce a short status summary for the interface.

How much usage can agent summaries consume?

The share varies by workflow. Better Stack measured about 8% in a realistic multi-agent build and 12% in a synthetic test; other reports vary.

Can I turn off agent summaries in Claude Code?

There is no dedicated native toggle described in the tests. Setting fork subagents to zero stopped summaries, but removes the parent-context inheritance those agents provide.

Is it safe to install a script that blocks the summary timer?

Treat runtime hooks cautiously: inspect the code, avoid blindly piping remote scripts into a shell, and expect updates to break the workaround.