Coding agent skills & methodology

• addyosmani/agent-skills packages senior-engineer workflows (define→plan→build→verify→review→ship) as reusable skills so agents follow consistent quality gates across a feature's lifecycle.
• obra/superpowers is a full agentic dev methodology built on composable skills, cross-compatible with Claude Code, Codex, Cursor, Copilot CLI and more.
• tirth8205/code-review-graph builds a Tree-sitter structural map of a repo with incremental updates so agents stop re-reading whole codebases on review tasks; benchmarked context reductions included.
• ayghri/i-have-adhd is a 10-rule skill that forces agents to lead with the action, number steps, and stop burying the answer, useful pairing for your caveman/lead-with-recommendation setup.
• zhaoxuya520/reverse-skill auto-routes pentesting/reverse-engineering agents (APK, binary, CTF, frontend crypto) to the right on-demand toolchain across Claude Code, Kiro, Cursor, Cline.

Claude Code & Codex releases

• Claude Code v2.1.224 adds self-hosted-runner (run web/mobile/desktop sessions on your own machines) and an archive plugin source for installing from a zip without git/npm.
• Claude Code v2.1.218 moves /code-review to a background subagent so review no longer fills the main conversation.
• Claude Code v2.1.215 stops auto-running /verify and /code-review, now opt-in only, matches recent Anthropic push toward explicit invocation over silent automation.
• Codex CLI 0.147.0 adds portable Agent Plugins with search across local, personal, workspace, and remote plugin catalogs.

Models

• GLM-5.2 marks a capability threshold for open agentic models per Nathan Lambert's tracking, worth watching if you want a non-Anthropic terminal-agent baseline.
• Claude Fable 5 and new AI safety fables covers Anthropic's Fable release and a walkback on silently manipulating AI-research queries in favor of a transparent classifier.
• Kimi K3 is now available in GitHub Copilot, billed at provider list pricing under usage-based billing, another open-weight model landing inside a mainstream coding agent.
• lyogavin/airllm runs 70B models on a 4GB GPU (and Kimi K3's 2.8T MoE under 4GB) via layer-wise loading, no quantization needed, relevant if you self-host anything.
• Inside vLLM: Anatomy of a High-Throughput LLM Inference System breaks down continuous batching, prefix caching, and specdec; HN consensus says the real wins were batching/KV-cache abstractions, not the marketed paged attention.

Agent memory & state

• TencentCloud/TencentDB-Agent-Memory is a team-level memory hub turning chats/docs/code into four shared, governed memory assets (Chat, Skill, LLM-Wiki, Code-Graph) usable across agents and frameworks.
• SnapState offers persistent state for AI agent workflows, a lighter-weight alternative angle on the same problem TencentDB is solving at team scale.

Agentic SaaS

• CopilotKit Channels SDK brings any AG-UI-compatible agent natively into Slack/Teams with interactive UI; SDK client is MIT but the backend runner is proprietary and license-gated.
• Cloudflare OS pitches itself as an "AI operating system" for companies built on Cloudflare's edge platform.
• Ticketdesk AI is another entrant in AI-agent customer support tooling.

Misc

• Humans missed 1 in 3 threats approving AI agent commands across 40k game runs quantifies human-in-the-loop failure rate on agent command approval; community take: permission prompts are legal cover, not real security, sandbox with file-level controls instead.
• Muse Code is Meta's terminal agent for long-horizon coding tasks.
• Superlog Responder is a free, open-source AI bug-fixing agent.
• embabel/embabel-agent is a JVM (Kotlin/Java) agent framework mixing LLM prompts with code and domain models, with goal-directed path finding.
• Run Claude & Codex in Your Browser thread on browser-based access patterns for both agents.