Coding agents & harnesses

• addyosmani/agent-skills: packaged define→plan→build→verify→review→ship skill set for coding agents, aims to encode senior-engineer workflow discipline as reusable agent instructions.
• React for Agents: Flue gets hooks: Astro creator Fred Schott ports React's hooks model into his agent meta-harness, signalling harness design (not raw model capability) as the next agent-framework battleground.
• Codex: 0 to 10M users: OpenAI's Akshay Nathan on Codex MAU up >10x since January, framed as an agentic-interface play for the ~100x people who "use code" but can't write it.
• Running three Claude Code sessions in parallel with git worktrees: community writeup on the worktree-per-agent pattern for parallel Claude Code sessions.
• ClaudeMax: free open-source Claude Code cost analyzer plus a public usage leaderboard, brew install claudemax.
• Fleet view for multiple Claude/Codex sessions: tmux-free dashboard for watching many concurrent coding-agent sessions at once.
• nopus: deterministic detector/rewriter for overly complex model responses, aimed at trimming Opus-style verbosity.
• Claude Code plugin turns agents into pixel creatures: novelty visualization layer that renders running agents as watchable sprites.

Multi-agent systems & security

• Patterns and problems in emerging multi-agent systems: Anthropic's Frontier Red Team study on agent-to-agent interaction risk as agents increasingly share codebases and markets; community notes Opus 5's shift toward agent-legible (less human-readable) design as a deliberate bet on designed coordination over raw intelligence.
• MCP reaches a security inflection point: 21,000+ internet-facing MCP servers found exposed, 92% without OAuth; OWASP MCP Top 10 forming, community pushing for npm-style signed lockfiles over server code before trusting it.
• ProofRun: local, cryptographic verification receipts proving which checks actually ran against the current code state, direct answer to agents claiming "tests pass" from stale memory.

Agent memory & context

• TencentDB Agent Memory: team-level shared memory hub turning chats/docs/code into four reusable asset types (Chat Memory, Skill, LLM-Wiki, Code-Graph) governed across agents and frameworks.
• SnapState: persistent state layer for AI agent workflows, aimed at surviving context resets across long agent runs.
• Ontologies Are So Back: argues formal ontologies are being revived to keep probabilistic agents inside deterministic boundaries, a structured alternative to pure prompt-based context management.

Models & training

• GLM-5.2 is the step change for open agents: flags GLM-5.2 crossing a capability threshold for open-weight agentic models, paired with a new open RL recipe paper for terminal agents.
• GLM-5.3: follow-on release billed as a coding capability jump from scaled post-training on the same base model.
• Frontier post-training recipe review: podcast with Finbarr Timbers on what it'd take to push an Olmo-style open recipe to frontier post-training quality.
• Soup: one-YAML LLM fine-tuning CLI with layer streaming, claims an 8B model trainable on a 4GB laptop GPU.
• Inferock Bench: independent, verifiable receipt system for every LLM API call, adjacent to the ProofRun trust-verification trend.

Agentic SaaS & automation

• citrolabs/ego-lite: browser built to share your logged-in session state with coding agents (Codex, Claude Code) without disrupting your own tabs, positioned against browser-use/agent-browser's separate-browser model.
• stablyai/orca: desktop/mobile ADE for running a fleet of agents (Codex, Claude Code, OpenCode, Pi) side by side, each in its own worktree with usage/rate-limit tracking.
• BrowserAct Cloud: one-prompt web scraping product, agent-driven data extraction as a hosted service.
• Automated Outlook/Zoom scheduling from email content: agent pipeline that schedules meetings by parsing who-said-what in email threads.
• HKUDS/CLI-Anything: generates CLI wrappers for arbitrary software so agents can drive it, plus a CLI-Hub package registry for sharing agent-usable CLIs.
• Phind: GPT-4-powered developer search; community flags it eliminates SEO junk but hallucinates on sparse results and has an unsustainable unlimited-API-cost model.