A status line of three to five words is eating 16% of spend? Claude Code's invisible calls

Hi, it's Haru. This week's story will make anyone running a lot of agents side by side do a double take. Remember the short status line in Claude Code's agent panel, the three-to-five-word one that says something like "Fixing tests"? Users digging through their own traces found it was burning far more tokens (the units you pay AI providers by) behind the scenes than anyone expected. I didn't run this myself. Instead I went through the Gist, the Hacker News thread, the GitHub issues and a third-party verification write-up. Here's what I found.

The full transcript is resent every 30 seconds, and three to five words come back

It started with an analysis Flamerix AI published as a GitHub Gist. Claude Code can export logs through OpenTelemetry, the standard framework for emitting telemetry from software, and Flamerix AI grouped those logs by query_source, a label that says what each call is for. The breakdown showed that in interactive sessions with background agents, Claude Code was resending each agent's entire transcript to the model roughly every 30 seconds. All that came back was the three to five words shown in the panel.

The numbers are striking. Of a week's estimated spend, these agent_summary calls made up 16.4%, more than the main conversation (the part where you actually talk to the AI) at 15.7%. By API requests the share was 19.7%, roughly one call in five. The input-to-output ratio was about 9,000:1: the model reads a huge amount and is asked for a single phrase. On top of that, 96% of the summary spend went to Opus, one of Claude's most expensive models. It's like making a top restaurant's chef read the whole daily log every 30 seconds just to ask, "What are you doing right now?"

The feature has been on by default since v2.1.232 (August 2026), and the measurements covered versions 2.1.257 through 2.1.312. Its share of weekly spend grew from 3.5–7% in July to 16% in October. Measurements from four other people collected by Flamerix AI came in at 18–26%. Flamerix AI also released no-agent-summary.js, a Bun preload hook (a small script loaded at startup that swaps out behavior) that stops the calls without modifying the Claude Code binary.

Everyone started checking their own bills. Some saw 16%, others 3%

The Gist drew 558 points and 225 comments on Hacker News. The interesting part was how quickly people began posting their own numbers. simon83 converted their week's usage to API pricing and wrote that of $1,970 in total, about $380 (16%) went to the hidden status-line calls. That's a real slice of the bill spent on a few words in a panel.

Not everyone had the same experience, though. jake_kozaur saw "about 8% at most," and _boffin_ saw "about 3% across the whole team." The same feature ranges from 3% to more than 20%. I think that is the most important point in this whole story.

More detailed measurements are in issue #86879 on anthropics/claude-code, opened on August 15. rflcrz examined a 1,496-request session that ran 27 subagents. agent_summary accounted for 339 calls (22.7%), 22.2% of raw tokens (62.7 million tokens), and 18.0% of weighted cost. Each call produced only about 65 output tokens. According to the issue, a single agent with a 380,000-token context generates about 45 million cache-read tokens per hour from the status line alone. (Cache reads are counted when previously sent content is reused; they're discounted, but not free.) The interval is fixed at 30,000ms. Other users reported that with 10 or more agents, the calls made up 14–26% of requests and 10–22% of weighted token spend.

This was the part that alarmed me most. The calls are marked skipTranscript (excluded from the conversation log), so they don't show up in cost-tracking tools either, and actual consumption looks 10–26% lower than it is. In one Opus 5.5 example, visible cache reads came to 1.184 billion tokens, while invisible ones came to 338 million. The harness (the layer that wraps the model and connects it to tools and the interface) makes calls that appear on your bill but not on your own dashboard. That is exactly the kind of blind spot agent observability is meant to remove.

Anthropic's answer: about 2% overall, but "it's a lot"

In the same Hacker News thread, edwinarbus from the Claude Code team responded. Across all Claude Code users, he said, agent_summary accounts for about 2% of cost at API list prices and just under 4% of requests, rising to about 6% for terminal use alone. He said the 16.4% figure reflects "a team running almost exclusively background agents, not a typical session."

He still acknowledged plainly that he agreed it is "a lot for a status line." A fix for cases where the cache misses shipped in v2.1.311 and later. The team is now building a much cheaper approach that sends a short note to a smaller model. In other words, instead of having Opus read the full transcript, only the key points would go to a lighter model. I thought this was a genuinely good response: it pushed back with data without denying the problem.

The 2% and 16% figures don't actually contradict each other. The overall average includes the large majority who run only one agent. Look only at the minority running fleets of agents and the figure rises to 16–22%. A small average doesn't mean your own bill is small. Then again, if you aren't at 16%, there isn't much to worry about. Either way, the only way to know is to look at your own traces.

Can you turn it off? The unofficial hook works, but there's no official setting yet

The workaround in circulation is Flamerix AI's no-agent-summary.js. Third-party outlet yoDEV reportedly tested it on macOS 27.0.1 (Apple Silicon) with Claude Code 2.1.312, using three background sessions. Counting [AgentSummary] Forking entries in the debug log, it saw five calls without the hook, five in pass mode (a pass-through setting used for comparison), and zero in off mode. That confirms the calls go away. But yoDEV states explicitly that it did not verify the 16.4% cost figure. The calls stop, but no third party has yet confirmed how much that saves.

The hook is also unofficial and works by intercepting Bun's timers, the mechanism that fires tasks at set intervals. Every time Claude Code updates, you need to check that it still works, and an issue for exactly that has reportedly been opened on Flip-Engineering/baton. For now, the only official option is CLAUDE_CODE_FORK_SUBAGENT=0, but that also gives up the benefits of forking subagents (branching them off with a copy of the parent's context). You want to stop just the status line, and you end up losing a useful agent feature along with it.

Meanwhile, #86879 has been open for more than two months since August. User distrust runs deep, partly because an earlier issue about cache expiry (TTL), #46829, was closed as not planned and described as "an intended optimization." The thread also featured plenty of criticism that the harness has become too bloated, along with the view that "the company selling the model and the company building the harness that calls it should be separate." The same company decides how often the calls happen and collects the fees for them. I can see where that feeling comes from.

Haru's takeaway: look at your own traces, not the average

explainx reportedly lined up the figures so far: Flamerix's 16.4%, #86879's 22.2% of raw tokens and 18.0% weighted, and Hacker News's "about 8% at most." Its conclusion was that the impact is real but varies widely with usage. Four factors matter: the number of agents running at once, context length, agents left idle but still running, and expensive models such as Opus. For normal use within a single session, it said, the impact is small.

What I found most interesting is that users found this, not Anthropic. Simply splitting the OpenTelemetry output by query_source surfaced the invisible calls. Harness engineering (designing the environment agents run in) is all the rage right now, and it seems users now also need to be able to count what their harness is calling behind the scenes.

What to do is simple. If you run many agents side by side, first work out agent_summary's share in your own traces. If it's above 10%, try the workaround hook or close agents you aren't using. If you install the hook, check the debug log after every update to confirm Forking stays at zero. If it's only a few percent, just wait for Anthropic's cheaper version. There's no need to panic over a "status line eats 16%" headline, but if you haven't measured, you don't yet know which group you're in.