Coding agents & CLI orchestration

• Claude Code next version: subagents run background by default, keep chatting while they work, foreground opt-in via explicit ask (tweet).
• Orca — ADE running Codex/ClaudeCode/OpenCode/Pi side-by-side, each own worktree, one dashboard, usage/rate-limit tracking across providers.
• herdr — terminal agent multiplexer, detach/reattach over ssh, sessions survive restarts, socket API lets agents spawn/control panes themselves.
• jcode — new coding agent harness targeting multi-session workflows and performance ceiling.
• pi — self-extensible agent harness: unified LLM API, agent loop, TUI, CLI, modular npm packages.
• MagicPath 2.0 — multiplayer canvas where humans + Codex/Claude Code build prototypes together, live agent view.
• mattpocock/skills — small composable Claude/agent skills library, model-agnostic, alternative to heavier frameworks (GSD/BMAD/Spec-Kit).
• Vercel's Andrew Qu on agents as new software category: eve framework, skills/sandboxes/agent-readable sites as the emerging stack (latent.space).
• Autoresearch: self-improving agent "recipes" and outer feedback loops, humans still central (latent.space).

Agent security & infra

• agent-security — open-sourced preflight gate for coding agents: catches poisoned repos, hostile content leaking into commits, destructive git commands before trust boundary crossed.
• cua — open computer-use drivers for cross-OS agent fleets, drives native apps in background without stealing cursor/focus, plus benchmarks (Cua Bench).
• SnapState — persistent state layer for agent workflows (HN).
• Axon self-documents its API for agent-to-agent hiring/payment, no human in loop for integration (tweet).

Token/context economics

• Deep research pipeline burned a whole Claude Max 5x limit in 30 min; lesson: dynamic prompts kill caching, static retrieval beats micro-optimization (quesma.com).
• Reddit skill routes "thinking" to Claude, grunt coding to cheaper/free models as executor, judge/executor split for cost control (reddit).

Open model landscape

• GLM-5.2 flagged as step-change for open agents, capability threshold worth tracking (interconnects.ai).
• Open model bonanza: Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 all shipped same month, CAISI ran independent V4 assessment (interconnects.ai).
• Claimed roadmap chatter: GLM-5.2 near-Opus, Kimi-K3 near-Fable, Qwen-3.8 (2.4T) and DeepSeek V4 GA ($0.0028/M) expected to beat Opus (tweet, unverified vendor claims).
• Alibaba Qoder claims #1 China AI coding market, 47.6% share per IDC, 5M+ users (tweet).
• LoRA Speedrun — public wall-clock leaderboard for fine-tuning Qwen2.5-1.5B to GSM8K threshold, independently re-run 3x on fixed hardware, modded-nanogpt for fine-tuning.

Vibe-coding & agentic SaaS

• Founder claims vibe-coded custom CRM replaced Salesforce, cut $600k/yr software spend to zero, better AI-agent integration than off-shelf (tweet, anecdotal).
• Vibe-Trading — personal trading agent toolkit, note: security warning re impersonating X/token accounts.
• WrenAI — open-source GenBI, agents generate/govern text-to-SQL dashboards across 20+ data sources (BigQuery, Snowflake, Databricks etc) via governed context layer.
• AstrBot — open-source agent chatbot platform integrating IM platforms + LLMs + plugins, positions as openclaw alternative.

CLI/TUI

• Homebrew 6.0.0 — new tap-trust security model, faster internal JSON API, Linux sandboxing, macOS 27 support; community noting forced-upgrade friction pushing users toward Mise for version pinning.