Coding agents & harnesses

• Headlong: open-source microharness (<10K lines bash) gives agents a persistent self-guided inner-monologue loop instead of reactive task-then-stop; community says custom harnesses keep beating LangGraph-style frameworks on integration.
• Phoenix: new coding agent purpose-built for iOS/macOS app dev.
• Dropstone: "AI runtime that remembers, learns, and acts everywhere," another persistent-memory agent runtime launch.
• openclaw: cross-OS personal AI assistant gateway wiring models, tools, messaging channels, and companion apps behind one operator-owned hub.
• Agents write fast, verification is where we customized our Claude Code workflow: practitioner thread on shifting Claude Code workflow investment from generation to verification, the actual bottleneck once agents write fast.
• If your digital product shop is profitable, which Claude Code workflows earned their keep?: crowd-sourced field data on which Claude Code workflows pay off vs waste time in real shops.
• ai-job-search: Claude Code framework for job hunting, evaluate postings, tailor CVs, prep interviews, useful pattern for domain-specific agent pipelines.
• ai-engineering-from-scratch: community curriculum for building/shipping AI engineering skills, i18n'd repo.

LLM mechanics, memory & tooling

• llm-anthropic 0.27: Simon Willison's Anthropic plugin for llm updates for compatibility with anthropic-sdk-python v1.0.0.
• Everyone is building more memory for Claude Code. After 1,000 lines of CLAUDE.md, I'm convinced it needs a delete key: contrarian take, memory accretion is the problem, not the fix, pruning matters more than accumulation.
• claude-obsidian: self-organizing Obsidian knowledge graph built on Claude Code skills, Karpathy LLM-wiki pattern, plain-Markdown ownership.
• freellmapi: aggregates 34 free-tier LLM providers (635 endpoints, 7.4B tokens/mo) behind one OpenAI-compatible endpoint with smart routing/failover, handy for cheap experimentation.
• modlens: vision plugin bolting image understanding (OCR/layout/semantic JSON) onto text-only coding harnesses like DeepSeek/GLM.
• LLMs could control their host machines by exploiting inference engines: threat model for a malicious model escaping the harness into the GPU host itself; community pushback says sandboxing belongs at the VM/container layer, not the harness.

Vibe-coding & repos

• Open source tool to back up Claude Code memories, agents, commands, and skills to a private GitHub repo: version-control your Claude Code config the same way as code.

Agentic SaaS

• Decawork: control panel for internal company AI agents and tools.
• Offloop: shared workspace where humans and AI agents co-execute work.
• Trama: plain-language macOS automation builder.
• Cortex by SKYNETLAB: memory layer that decides what's worth persisting for agents.
• Lucid Train: system-design generation for new and existing codebases.
• Navigara: ties AI spend directly to product roadmap for cost accountability.

CLI/TUI

• Treebar: single-pane view of every active git worktree, aimed at multi-agent/Codex sprawl.
• What's new in Emacs 31.1: no headline feature, but built-in treesitter/eglot equivalents cut third-party config burden.
• Show HN: A Modern GUI Library for Ada: CSS-styled, XML-layout GUI toolkit on SDL3; community flags Ada/XML choices as dated vs Rust/Dioxus alternatives.