Coding agents & CLI

• Claude Code ships v2.1.261, track changelog for skill/tool-use changes.
• Codex CLI cuts two releases same window, 0.153.3 and 0.153.4, fast iteration on the Rust rewrite.
• anomalyco/opencode: new open-source coding agent, another entrant in the Claude Code/Codex/Aider space worth benchmarking.
• GitHub Copilot's weekly release notes, Aug 31 round up incremental agent/editor changes.
• GitHub's Copilot app guide for beginners covers managing agent work from the standalone app, signals Copilot pushing a task-management surface beyond the editor.

Frontier models

• GPT-6 Astra launches and is already generally available in GitHub Copilot, major model swap for anyone routing Copilot agent tasks.
• CodeRabbit's eval of GPT-6 Astra in code review weighs accuracy gains against privacy and cost, relevant when picking a review-tier model.
• GitHub's Project HydraFusion routes across multiple frontier models per request for best-of quality, concrete multi-model orchestration pattern.
• Google Gemini 3.8 Flash and Cyber ship, Flash tier keeps getting cheaper/faster for high-volume agent calls.
• Z.ai's Jie Tang on GLM 5.3 and post-training scaling laws: argues param count matters less than post-training compute, worth reading for where capability gains now come from.
• radixark/miles: new enterprise RL framework for LLM/VLM post-training, forked from slime, an option for custom fine-tune/RLHF pipelines.

Token efficiency & Claude Code tricks

• Portal by Spotify claims a 90% Claude Code token cut, read for the caching/context technique behind it.
• JuliusBrussee/caveman packages "talk like a caveman" compression as a Claude Code skill, claims 65% token reduction, same trick already running here.
• Reddit writeup on a no-secrets, restricted-network Claude Code setup against real production infra, concrete sandboxing/permissions reference.
• Someone's months-long fight to stop a Claude Code pipeline grading its own homework, then found their own published plugin had the same self-evaluation flaw, cautionary tale on LLM-judge bias.
• Reddit thread on parallel vs sequential agent systems surveys what research actually says about multi-agent architecture choices.

Agentic SaaS & vibe-coding

• Swipe: AgentsView and search the work your agents already did both chase the same gap, making past agent runs queryable and observable.
• Swipe: Graft, another agentic dev tool entry.
• Swipe: sketch an app, let AI build it on the same page, sketch-to-code vibe-coding pattern.
• Swipe: let an AI agent make your first video rough cut, agentic workflows extending past code into media editing.

Also shipped

• Omarchy and Clockwork launched on Product Hunt, worth a skim if either overlaps the stack.