Coding agents & LLM mechanics

• Claude Code v2.1.247 adds SendFeedback tool (drafts /feedback reports) plus configurable spinner tips.
• Codex 0.150.0 adds @-mention cross-task references (read/create/message other Codex tasks from terminal); 0.150.1 fixes remote compaction to count retained images toward token budget, trimming old images first.
• Qwen3.8-Flash-Next: open-weights MoE, 125B total/6B active params, early preview of Qwen4 architecture, big speed win from sparse activation.
• GLM-5.3-Flash: Zhipu's cheap flash model, $0.15/$0.50 per 1M in/out, undercutting most flash-tier competitors.
• Anthropic's AI-native SDLC playbook: six-stage dev lifecycle where line-by-line review gets replaced by something else, worth checking what.
• Reddit thread on which Claude tasks justify the most capable model, model-routing tradeoffs from practitioners, same territory as cost-aware pipeline routing.
• Reddit thread on what belongs in the agent binary vs. the system prompt, design split for agent tooling vs. prompt-encoded behavior.
• Reddit workflow for using third-party subagents in Claude Code.
• GitHub Copilot's global model policy now enforced org-wide on Business/Enterprise plans, admins control which models devs can pick.
• loveholidays case study: using Codex to push software dev access beyond engineering into the wider business.
• Simon Willison quotes Paul Dix on AI writing 1M LOC then refining it into reliable software running on millions of machines, notable because there was no oracle for correctness this time.
• acceptmarkdown.com: serve Markdown via Accept: text/markdown content negotiation so agents skip nav/scripts/DOM and read prose directly, cuts context waste, needs top chatbots on board to matter.

Vibe-coding & agent skills

• browser-use/browser-use: lets agents (Claude Code, Codex, Cursor) drive a real browser, click/type/fill forms from a task description, one-shot setup prompt included.
• K-Dense-AI/scientific-agent-skills: 163 validated skills for scientific research (bio/chem/med/drug discovery), now framework-agnostic under the open Agent Skills standard, used by 175k scientists.
• ConardLi/garden-skills: curated Agent Skills collection (web design, image gen, knowledge retrieval) for Claude Code/Cursor/Codex.

Agentic SaaS

• Lovable CTO on the future of SaaS being apps agents use, not humans, pitching a "digital brain" connecting daily tools instead of conventional app UIs.
• DeployHermes: persistent agent hires with roles, memory, and skills.
• Knack MCP Server: HIPAA-compliant backend for AI builders.
• MCP-Builder.ai: fast path to wire your data into MCP tools.
• OpenComputer: "Firebase for Agents", backend-as-a-service framed for agent workloads.
• Evidence Core: live analytics dashboards built via coding agents.
• Arena Agent Mode: GitHub-integrated agent mode taking ideas to shipped PRs.

CLI/TUI & build tooling

• Tailcat: netcat-like tool riding Tailscale's WireGuard data plane without needing Tailscale's control plane, community flags vendor lock-in vs. open alternatives like netbird.
• mold: A Massively Parallel Linker: arXiv paper on the parallel linker design, community notes Wild benchmarks faster and mold's Linux-only stance blocks enterprise adoption.