First screen, captured 1 day ago. Hover to scroll the page, or open the video.

Claude Code regularly sends each background agent’s transcript to its model

What HN thinks

A status line that costs more than the conversation it's summarising, and a thread that can't decide if that's greed or a bug.

Token waste reflects misaligned incentives
Many argue a model vendor building its own harness has little reason to be frugal with tokens, especially for enterprise customers billed at API rates, and some suspect IPO-driven usage padding.
Bloat or bug, not malice
Others think Anthropic engineers simply don't feel token costs and keep piling features into a bloated harness; cached reads make each call cheap but they add up.
Switch to leaner harnesses
A sizable group recommends alternatives like Pi or oh-my-pi, or patching or pinning old Claude Code versions to avoid background overhead.

Consensus: Most agree a token-burning summary call should at least be user-toggleable and documented; an Anthropic engineer said it averages around 2–6% of cost and that cheaper fixes are coming, which many found unsatisfying.

Recommended in the comments

Pi (tool): Praised as lean, customizable and more token-efficient than Claude Code.
oh-my-pi (tool): Called the best harness by a comfortable margin and compatible with existing Claude Code setups.
codex-thread-bridge ↗ (tool): Used to offload work from Opus to cheaper models and reduce Anthropic quota burn.
ccusage (tool): Used to measure API-equivalent spend against subscription plans.

Summarised by Claude from the comment thread 12 hours ago.

Points: 498
Comments: 210
Best rank: #3
Posted: 2026-10-29 by NikZaal
Category: Ai ml

claude-code anthropic bun token-usage opentelemetry llm-agents

All captures:

Full page

Captured 1 day ago. Click for the whole screenshot.

Looks similar

Screenshot of github.com: Show HN: Docx-CLI: agents read/edit Word docs using 1/2 the time and tokens

Show HN: Docx-CLI: agents read/edit Word docs using 1/2 the time and tokens (github.com)

53 points by kirillklimuk 3 months ago | 23 comments

Screenshot of github.com: Easy to use OpenID Connect client and server library written for Go

Easy to use OpenID Connect client and server library written for Go (github.com)

143 points by mffap 2 years ago | 40 comments

Screenshot of github.com: Show HN: My Claude quota ran out in 10 minutes, so I made a tool to find out why

Show HN: My Claude quota ran out in 10 minutes, so I made a tool to find out why (github.com)

47 points by sachinneravath 2 months ago | 39 comments

Screenshot of github.com: Cloudflare OS – an open-source AI productivity environment

Cloudflare OS – an open-source AI productivity environment (github.com)

21 points by berlianta 2 months ago | 3 comments

Screenshot of github.com: An SLM trained on $8 ESP32-S3

An SLM trained on $8 ESP32-S3 (github.com)

55 points by pavelai 2 months ago | 17 comments

Screenshot of github.com: Show HN: Plimsoll, a self-hosted sandbox for code your AI agent writes

Show HN: Plimsoll, a self-hosted sandbox for code your AI agent writes (github.com)

11 points by cagster 18 days ago | 2 comments

GitHub repository for Plimsoll, a self-hosted sandbox service for running untrusted AI agent-authored code that reports the isolation boundary used for each run.

ai-agents code-execution sandboxing self-hosted

Full page

Looks similar

Show HN: Docx-CLI: agents read/edit Word docs using 1/2 the time and tokens (github.com)

Easy to use OpenID Connect client and server library written for Go (github.com)

Show HN: My Claude quota ran out in 10 minutes, so I made a tool to find out why (github.com)

Cloudflare OS – an open-source AI productivity environment (github.com)

An SLM trained on $8 ESP32-S3 (github.com)

Show HN: Plimsoll, a self-hosted sandbox for code your AI agent writes (github.com)

Related stories

90% of Claude-linked output going to GitHub repos w <2 stars (claudescode.dev)

Show HN: A Claude Code skill to analyze your chess games (github.com)

Issue: Claude Code is unusable for complex engineering tasks with Feb updates (github.com)

Claude Code now supports hooks (docs.anthropic.com)

Claude Code can debug low-level cryptography (words.filippo.io)

Show HN: I wrote a RDBMS (SQLite clone) from scratch in pure Python (github.com)