---
title: 'In the News: October 29, 2026 (Morning)'
description: 'A team says Claude Code background-agent status summaries took 16.4% of its estimated spend. An Anthropic engineer puts the all-user figure near 2%.'
canonical_url: 'https://darkfactory.dev/news/2026-10-29-morning'
markdown_url: 'https://darkfactory.dev/news/2026-10-29-morning.md'
collection: news
date_published: '2026-10-29T07:05:00-04:00'
date_modified: '2026-10-29T07:05:00-04:00'
---

# In the News: October 29, 2026 (Morning)


One item cleared the bar this cycle: a cost report on the side calls a coding-agent harness makes for itself, with a vendor reply that disputes how typical the numbers are.

## 1. A telemetry report says Claude Code's background-agent status line cost one team 16.4% of its spend; Anthropic says 2% across users

**[Claude Code burns ~15-20% of your tokens on a 3-word status line: a Bun preload that turns agent_summary off](https://gist.github.com/flamerix-ai/c3905705cc4205e5de3011fa2e74efce)** · Flamerix AI · GitHub gist, October 2026 · [Hacker News thread](https://news.ycombinator.com/item?id=50284255)

Flamerix AI reports that Claude Code sends each active background agent's full transcript to that agent's own model about every 30 seconds and asks it to "describe your most recent action in 3-5 words". The answer only updates a status line. In one week of OpenTelemetry data from the authors' setup, these `agent_summary` requests made up 16.4% of estimated spend, against 15.7% for the main conversation, and 19.7% of API requests. They report about 9,000 input tokens per output token, with 96% of that spend on Opus. The spend figure is Claude Code's own per-request estimate, not an invoice, and the authors say roughly 98% of the calls read the transcript from cache.

Edwin Arbus, who says he works on Claude Code, replied in the thread that the calls only apply to background agents that are actively working, and that 16.4% looks like a team running mostly background agents. He gave Anthropic's figures for the previous week across users who send telemetry: about 2% of cost at API list prices, just under 4% of requests, and about 6% of cost in the terminal. He also said a fix for summaries that miss the prompt cache shipped in 2.1.311 and a cheaper mode is in progress. The gist's authors say a short-notes mode already ships, switched off, in 2.1.310.

The authors also publish a Bun preload that stops the summary timer. They tested it on Linux across Claude Code 2.1.257 to 2.1.312; macOS and Windows are untested. It matches the timer by source text, so an update can silently turn the summaries back on, and the agent status panel stops updating while it runs. They warn that changing a client's runtime behavior may conflict with a service's terms.

**Why it matters:** If you run many background agents, the harness's own side calls are a cost line your session logs may not show. The authors say these requests do not appear in the on-disk session logs, only in OpenTelemetry export, so a cost tool that reads transcripts would miss them. The two share figures differ by a factor of eight and come from different workload mixes, so measure your own by grouping requests by `query_source` before deciding whether it matters to you. These are the vendor's and a single team's numbers; neither has been independently checked.