Friday, October 23, 2026Vol. 1 · No. 236 · 70 stories on fileCollected from around the web
Section
AI Coding & Agents
12 stories · AI agents, AI coding / agents, AI coding / code review, AI coding tools / Claude Code, Claude Code, Computer use / agents, Testing / AI coding
AI coding / agents
AI coding / agents — via @theo
Replying to argofowl, who asked how to juggle multiple accounts in the Claude desktop app without…
By Theo - t3.gg · @theo · Oct. 18, 2026
Replying to argofowl, who asked how to juggle multiple accounts in the Claude desktop app without logging out and losing the sidebar, Theo says 'T3 Code + cliproxy for now.' T3 Code handles multiple accounts as separate Claude provider instances, each with its own CLAUDE_CONFIG_DIR. CLIProxyAPI is a local proxy that pools several CLI OAuth logins (Claude, Codex, Gemini, Grok) behind OpenAI/Claude-compatible endpoints with round-robin. Caveat: Anthropic blocked third-party harnesses from using subscription limits on 2026-04-04, so subscription routing through the proxy may need extra usage.
Andromedus: 'if you're not using Grok Bot with T3Code you're not at the frontier.' Screenshot shows Grok Bot messaging a T3Code agent session (Claude Opus 5.5) to continue filing GitHub launch issues; the agent filed them and linked them under a tracking issue. A T3Code MCP for Grok Bot is coming soon.
Claude writes ~90% of Claude Code's code. June–Sept merged PRs more than doubled while the share reverted fell 40%+; Anthropic blog explains how they changed code review and how to adopt it.
typecraft: 'ok, installing hermes! lets see how this works.' Short screen-recording video of the install; the post doesn't say which Hermes (likely Nous Research's Hermes Agent, unconfirmed).
Burned 76% of a weekly Claude limit in under 24 hours in one session orchestrating 5-7 parallel…
By Haz · @diegohaz · Oct. 9, 2026
Burned 76% of a weekly Claude limit in under 24 hours in one session orchestrating 5-7 parallel subagents (Opus 5.5 / Sonnet 5.5). Cause: subagent cache lives only 5 minutes vs 1 hour for the main agent, so each time an idle subagent got its next instruction it re-processed its whole history (often 500k+, up to 1M tokens) at ~25x the cached cost.
Fix: tell the orchestrator explicitly to spawn background threads (claude agents) and talk to them instead of using subagents; it defaults back to subagents unless you're very specific. Takeaway: use subagents sparingly for quick one-off exploration, never for work that may sit or run past 5 minutes. Asks Anthropic to fix subagents, warn, or default to background sessions.
Setting Claude Code /autocompact (autoCompactWindow) to 400k instead of the 1M default would have saved…
By rohit · @rohit3a · Oct. 8, 2026
Setting Claude Code /autocompact (autoCompactWindow) to 400k instead of the 1M default would have saved ~29% of weekly usage without hurting quality. His 7-day replay: 600k saves 18% (30 compactions/wk), 400k 29% (82), 300k 36% (126), 200k 41% (232). Savings flatten below 300k; 300k to 200k nearly doubles compactions for ~5 more points. Replay is within ~5% of actual spend.
Add a prompt-cache countdown to the Claude Code status line
By Vox · @Voxyz_ai · Oct. 6, 2026
Add a prompt-cache countdown to the Claude Code status line: shows minutes left on the cache and, once cold, how many tokens the next message will re-cache. Docs notes: on a subscription within plan usage the main conversation cache lasts 1 hour (subagents/workflows 5 min); on API key or usage credits it's 5 min. Switching models (incl. opusplan plan-mode toggles) recomputes everything; editing CLAUDE.md mid-session doesn't break cache but doesn't apply until /clear, /compact, or restart; /rewind can reuse the existing cache while /compact builds a new one. Includes a ready prompt for Opus 5.5 to add the segment (needs Claude Code 2.1.251+, refreshInterval 30). Docs: https://code.claude.com/docs/en/prompt-caching
Uses Cloudflare's decision API for computer-use clicks
By ThePrimeagen · @ThePrimeagen · Oct. 3, 2026
Uses Cloudflare's decision API for computer-use clicks: take probabilities to a power, then probabilistically center, and it lands on the right icon. Example prompt "Click the monitor icon in the menu bar" hits the monitor in an Omarchy menu-bar screenshot (grid of cell scores, target cell highlighted).
/claude-api prompt-audit is now also /checkup prompt-audit
By Lydia Hallie · @lydiahallie · Sep. 25, 2026
/claude-api prompt-audit is now also /checkup prompt-audit. The old name sounded API-only but it always worked on Claude Code setups — audits CLAUDE.md, skills, and agents for instructions the model no longer needs.
Unit tests are mostly pointless now: AI bloats them, locks undesirable behavior, and 100% coverage slop still breaks production. Models already nail algorithmic correctness; real failures are second- and third-order system effects. Prefer E2E behavior over piles of unit tests.
