MetaPatchAI home
AI

Gist claims Claude Code keeps re-sending background agents' transcripts: here's what Anthropic's docs say

Analysis by Stephen Pierce · Published , 1:40 PM ET

Oct 30, 2026. A GitHub gist titled "Claude Code regularly sends each background agent's transcript to its model" hit the top of Hacker News this week (573 points). It claims Claude Code re-sends background agents' transcripts to a model every ~30 seconds just to write a short status line. What does that mean for your privacy and your usage limits? Below we check the claim against Anthropic's own Claude Code documentation, the primary source, and keep the parts Anthropic confirms separate from the parts it doesn't.

Primary sources: https://code.claude.com/docs/en/agent-view (Row summaries) and https://code.claude.com/docs/en/data-usage

What Anthropic's docs say

The agent view docs confirm that each background session's row has a one-line summary written by a Haiku-class model, so you can see what a session is doing without opening it. According to the docs:

  • While a session is working, the row text updates at most once every 15 seconds from the session's own recent output, without sending a model request.
  • The model writes a fresh summary when each turn ends and, during a long turn, rewrites it every few minutes.
  • Each end-of-turn summary and mid-turn rewrite is one short Haiku-class request through your normal provider, "billed and handled under the same data usage terms as the session itself."
  • On a third-party provider or gateway with no Haiku-class model configured, the request uses the session's main model. Setting ANTHROPIC_DEFAULT_HAIKU_MODEL chooses one.

So Anthropic does confirm that summary requests are sent based on session content. It describes them as small, periodic and covered by the same terms as the rest of your session.

What the gist claims (unconfirmed)

The gist, by Flamerix AI (not affiliated with Anthropic), says it measured Claude Code's OpenTelemetry export and found an agent_summary request roughly every 30 seconds per background agent, sending the full transcript, usually to Opus. It claims this was about 16% of spend and nearly 1 in 5 API requests, and that there's no setting to turn it off. It offers an unofficial preload script to suppress it.

These figures come from one vendor's own telemetry and a linked GitHub issue. We have not verified them, and they don't match Anthropic's docs on cadence or model (every few minutes, Haiku-class). The difference may come from versions or provider setup. We are not linking or recommending the unofficial script. Patching your client's behavior is at your own risk.

Is this a privacy problem?

Based on Anthropic's docs, summary requests go to the same provider as your session, under the same data terms. They don't open a new destination for your code. What applies, per the data-usage page:

  • Consumer (Free, Pro, Max): 5-year retention if you allow data use for model improvement, 30 days if you don't. Change it at claude.ai/settings/data-privacy-controls.
  • Commercial (Team, Enterprise, API): 30-day standard retention, with Zero Data Retention available to qualified Claude for Enterprise orgs.
  • Locally: transcripts are stored in plaintext under ~/.claude/projects/ for 30 days by default (cleanupPeriodDays).

The practical issue is cost and usage limits more than a new privacy exposure: each summary is another model request on your bill or plan.

Settings you can actually use

From Anthropic's docs:

  • ANTHROPIC_DEFAULT_HAIKU_MODEL: on gateways or third-party providers, point summaries at a small model instead of your main one.
  • DISABLE_TELEMETRY=1: turns off usage metrics, which the docs say never include code, prompts or file paths.
  • DISABLE_ERROR_REPORTING=1: turns off error reports, which are redacted before sending.
  • CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: disables non-essential traffic at once, including feedback surveys.
  • cleanupPeriodDays: shortens local transcript retention.

We found no documented setting that turns off row summaries alone. That fits the gist's claim, but Anthropic doesn't say so directly.

"Claude Code adds your email to the prompt"

A related HN thread makes this claim. We found no Anthropic documentation confirming or denying it, so we treat it as unconfirmed. The data-usage docs do say error reports redact email addresses and other personal info before sending.

What's confirmed / not confirmed

Confirmed (Anthropic docs, read Oct. 30): row summaries exist; they use a Haiku-class request at turn end and every few minutes during long turns; 15-second updates send no request; same provider and data terms; fallback to the main model on gateways without a Haiku model; the retention periods and opt-out variables above.

Not confirmed: the gist's 30-second cadence, full-transcript and Opus claims, and its cost-share figures; that no off switch exists; the "email in the prompt" claim. Anthropic has not commented on the gist as far as we found.

Related

Up: AI