Coding agents

• Claude Code — Anthropic's terminal-native agentic coding CLI; core reference repo for anyone building on top of it.
• v2.1.200 — AskUserQuestion no longer auto-continues by default (opt into idle timeout via /config), and default permission mode flips to "Manual" across CLI, VS Code, and JetBrains.
• v2.1.201 — follow-up patch release, changelog not rendered in feed.
• Claude Code via Microsoft Foundry — routing Claude Code through Foundry reportedly sidesteps the 5-hour/weekly usage caps.
• Stick shift for Claude Code, updated — community tool for manually gear-shifting model/mode mid-session gets an update.

Fable 5 & the model landscape

• Fable's judgement — Simon Willison relays a tip from the Claude Code team: let Fable (and Opus) use its own judgement on how to work rather than over-dictating steps.
• Categorising models got harder — with Sonnet 5 now behaving Opus-like and Fable/GLM 5.2 in the mix, the old Opus/Sonnet/Haiku tiering heuristic breaks down.
• Benchmark ranks Sonnet 5, GLM 5.2, Nemotron 3 Ultra above Fable 5 — pushback on a widely-cited leaderboard's credibility.
• 10x-cheaper Fable 5 usage framework — a 10-80-10 routing/loop-engineering/context-memory scheme claimed to cut Fable 5 quota burn.
• Fable + subagent swarm is a "farce" — Fable itself reportedly called its own swarm-of-subagents pattern ineffective.
• Sonnet5-med vs Opus4.8-low for science/light coding — practical model-routing question for lighter analytical workloads.
• When will you stop reading the code? — trust-threshold question as models improve, relevant to how much review agentic output still needs.

LLM mechanics: context, caching, geometry

• Context rot, compression, why models "forget" — understanding context-window degradation explains why large-context dumps silently lose fidelity.
• Chat caching performance improvements — reported speedups from prompt-cache changes.
• Dispersion loss vs embedding condensation — claims small-LM embedding collapse is a fixable geometry issue; note embedding anisotropy/collapse has been documented since 2017–19, so treat as incremental not novel.
• Jamesob's guide to running SOTA LLMs locally — hardware breakdown from $2k to $40k; caveat is real local SOTA compute runs closer to ~$400k, and Apple M-series unified memory beats $3k+ GPU rigs on value.

Vibe-coding & agentic tooling

• ZERO — new coding harness built from scratch in Go, positioned as leaner than existing agent harnesses.
• Vibe coding is 80% debugging, 20% building — thread questioning whether Max-tier usage is worth it given the debugging ratio.
• iffable — open-source hook to stop Fable quota being spent on low-value grunt work.
• Agentic coding notes from Galapagos Island — dan luu's field notes on AI coding agents; notes megabyte-scale context windows have eliminated the complex agentic-loop workarounds this genre used to assume were necessary.
• Osloq — AI agent that reproduces GitHub issues automatically.
• langchain-ai/openwiki — CLI that generates and auto-updates agent-facing codebase docs via a daily GitHub Action PR.
• MaxKB — open-source RAG agent platform with ingestion pipeline, MCP tool-use, and model-agnostic backend support.

Agentic SaaS

• Glaze by Raycast — chat-driven Mac app creation, another entrant in natural-language-to-app tooling.
• Skills to fix agent blind spots for Effect — argues a distilled "skill" covering what agents get wrong about the Effect library gets 80% of the value; notes evals on skills remain hard.