Coding agents & LLM mechanics

• Asana cleared 5 years of engineering work in 2 weeks with Codex: Codex replaced Asana's testing system in 2 weeks for ~$12K vs ~$6M staffing estimate, concrete agentic-coding ROI number to cite.
• GLM-5.2 is the step change for open agents: Nathan Lambert flags GLM-5.2 crossing a capability threshold for open-weight terminal agents, worth tracking vs closed frontier.
• Ontologies Are So Back: Why AI Agents Are Reviving the Semantic Web: agent builders rediscovering ontologies/knowledge graphs to keep probabilistic agents inside deterministic boundaries.
• Claude Code cross-session messaging via ListAgents/SendMessage: user observes Claude Code autonomously using its own agent-to-agent messaging primitives mid-workflow, no prompt.
• Two Claude desktop agents communicating: another sighting of spontaneous inter-agent chatter, same emergent-coordination theme.
• Claude Opus5 fails absurd AAA benchmark but ships a working Three.js game anyway: benchmark score misleading, functional output from empty repo is the real signal.
• Prompts to get token waste measured: community thread collecting prompts that surface context/token waste, relevant to caching/context-window discipline.

Vibe-coding & agent infra

• Qwen 3.8 Max (2.4T) and 27B, new open weights for coding: Qwen back with a massive 2.4T MoE plus a 27B sibling, open-weights coding contender.
• volcengine/OpenViking: open-source context DB unifying agent memory/RAG/skills as a viking:// virtual filesystem, agent browses with ls/tree/find instead of vector-DB queries.
• SnapState: persistent state layer for AI agent workflows, checkpoint/resume primitive.
• Redesigning your vibe coding project: an approach: practitioner writeup on restructuring a vibe-coded codebase once it outgrows throwaway status.
• Open Index: structured-context builder for smarter agents, PH launch.
• stablyai/orca: ADE for running Codex/Claude Code/OpenCode/Pi side-by-side each in its own worktree, one dashboard, quick-open across agents/files/commands.
• mukul975/Anthropic-Cybersecurity-Skills: 817 cybersecurity skills mapped to MITRE ATT&CK/ATLAS/D3FEND/NIST, works across Claude Code, Copilot, Codex CLI, Cursor, Gemini CLI.
• VibePod 0.20: one CLI driving multiple coding agents with switchable logins per run.
• chaitanyagiri/munder-difflin: free local multi-agent harness wrapping Claude Code/Codex/Antigravity/Grok/Kimi/Qwen/OpenCode, works within your own subscription hourly limits, keeps working while you're away.

Agentic SaaS

• Shepherd Terminal: persistent terminal running Codex and Claude side by side.
• ElevenLabs MCP in Claude: create/manage ElevenLabs voice agents directly from Claude chat via MCP.
• The Inference Engineering Masterclass, Baseten: Baseten (fresh $13B round) on inference infra at the peak of the open-weights debate, decacorn-tier infra context for anyone running self-hosted agents.
• Gauge: "agent-led growth", gets your product written into every customer's codebase.
• CrewTower: controls your agent fleet from the Mac notch, glanceable multi-agent status.
• Phind: GPT-4-powered web search for developers: dev-focused search engine, good for niche library discovery, weak on sparse-result hallucination.
• Superflow AI: AI agents QA your Webflow site pre-launch.

CLI/TUI

• Codex CLI 0.148.0: adds /export to dump TUI conversations to Markdown and codex exec fork to fork/archive/restore sessions from the resume picker, direct answer to session-management gaps vs Claude Code.
• Best site for browsing Claude Code Skills?: community thread hunting for a skills-discovery hub, gap in the ecosystem worth noting.