Coding agents

Anthropic shipped Claude Sonnet 5, its most agentic Sonnet yet, now the default in Claude Code with a 1M-token context window and $2/$10-per-Mtok promo pricing through Aug 31. Early takes call it close to Opus 4.8 at a fraction of the cost, but Anthropic's own post concedes Opus 4.8 is still cheaper and better on agentic/search/computer-use benchmarks, muddying the pitch. The efficiency story is contested: one thread says Sonnet 5 beats Opus by generating ~2x the tokens per task, a follow-up reframes that as OpenAI-style token efficiency being the real win Sonnet 5 lacks, and AINews attributes the gap to tokenizer changes plus 3-6x more turn-taking — Simon Willison's rundown of the official docs is the more actionable source than the launch post itself. Sonnet 5 also went GA in GitHub Copilot. Copilot Agent separately landed in JetBrains AI Assistant. Cursor shipped an iOS app for driving coding agents from a phone.

Fable/Mythos 5 access whiplash: export-control lift announced, then ban lifted, then usage-credit clawback announced for after July 7, netting one user 3 days full usage instead of the promised 15. Claude Desktop is now on Linux (Ubuntu/Debian, beta), giving Claude Code/Cowork/chat parity with browser and terminal.

LLM mechanics

Anthropic is formally naming "context engineering" — curating what the model sees — as the successor to prompt engineering, with bloated context framed as the silent killer of agent reliability. Adjacent buzzphrase "loop engineering," credited to Claude Code's Boris Cherny and OpenClaw's Peter Steinberger, is catching on as the term for the iterate-at-length loops that let agents build software end-to-end. Lilian Weng's deep dive on scaling laws reaffirms the loss-vs-compute power law as a reproducible predictor, not a one-off fit.

Vibe-coding & repos

obra/superpowers packages a full agentic-skills methodology — composable skills plus bootstrapping instructions — for coding agents, and is now hiring a community/code maintainer. One power-user reports running Claude Code entirely from a wiki as a cross-repo master controller with Todoist as task tracker, an unusual but instructive orchestration pattern.

Agentic SaaS / infra

google/agents-cli turns any coding assistant into a builder/deployer for agents on Google Cloud's Agent Platform, adding skills and commands for build/scale/govern/optimize workflows. diegosouzapw/OmniRoute is a free AI gateway aggregating 231+ providers behind one endpoint (Claude Code, Codex, Cursor, Cline, Copilot) with claimed 15-95% token savings via compression and auto-fallback — treat the "~1.6B free tokens/month" claim skeptically, it's vendor-reported.