Coding agents

Coding agents & CLI tooling

• Claude Code v2.1.211 adds --forward-subagent-text for subagent text/thinking in stream-json, plus fixes bidirectional-override/zero-width/look-alike char neutralization in permission previews relayed to chat channels (prompt-injection defense).
• Claude Code v2.1.221 ships VSCode Focus view (collapses tool activity into per-turn summary) and mode: "mask" for sandboxed credential files on Linux/WSL.
• Claude Code v2.1.212 reworks /fork to spawn a background session (own row in claude agents) instead of an in-session subagent; that role now lives in new /subtask.
• Claude Code v2.1.214 fixes real security bugs: single-segment dir/** allow-rules over-matching nested dirs, a Windows PowerShell 5.1 permission-check bypass, and Bash checks failing open instead of closed.
• Claude Code v2.1.215 stops auto-running /verify and /code-review, now opt-in only.
• GitHub Copilot cloud agent: reasoning level now configurable per task, and automations can trigger off issue/PR comments (e.g. "generate docs" via comment).
• GitHub Copilot: several models deprecate Sept 1 2026 across Chat, inline edits, agent mode, completions, plan migrations now.
• free-claude-code: proxy layer to run Claude Code/Codex/Pi against your own provider/local models instead of paid subscriptions, admin UI to validate providers.
• mpai: makes existing Codex/Claude Code sessions multiplayer.
• Reddit thread comparing Sonnet 5 high-effort vs Opus 5 low-effort, relevant to your reasoning-tier routing calls.

Agentic SaaS / orchestration

• Orca: ADE for running fleets of parallel coding agents (Codex/Claude Code/OpenCode/Pi), each in own worktree, one tracked view, desktop+mobile, uses your own subscriptions not per-seat billing.
• Hoplite (YC S26, HN launch): one-thread-one-VM cloud agent deploy, HN notes it's functionally identical to Amp's existing model.
• AgentSky: cloud-hosted agents, any harness + any LLM.
• SnapState: persistent state layer for AI agent workflows, addresses the "agent forgets everything between runs" problem.

LLM/model releases

• GLM-5.2: interconnects.ai calls it a genuine capability threshold for open agentic models, worth tracking vs closed frontier for agent workloads.
• Qwen 3.8 Max (2.4T) + 27B: Qwen back with a monster MoE flagship plus small model, signals renewed open-weights competition for coding/agent use.
• Cloudflare Workers AI: running Kimi K-series and GLM at scale, HN pushback that quantization isn't disclosed on the product page.
• Baseten inference engineering masterclass: infra deep-dive off a $13B round, includes a Kimi K3 breakdown, relevant if you care about inference cost/latency tradeoffs behind agent backends.

Vibe-coding / local inference

• Swiftlet (HN): runs 80B Qwen in 4.3GB RAM on Mac and 35B on iPhone by streaming MoE experts from disk on demand, HN flags prefill as the real bottleneck (~10 tok/hr disk-swap degenerate case).
• AirLLM: runs 70B+ models (up to Kimi K3 2.8T) on 4-12GB GPUs without quantization by streaming layers, useful if you want to run big local models on modest hardware.
• ds4 (DwarfStar): narrow, self-contained local inference engine tuned specifically for DeepSeek V4 Flash/GLM 5.2 on Metal/CUDA/ROCm, not a general GGUF runner.
• livekit/agents: framework for realtime voice AI agents (STT/LLM/TTS pipeline), server-side multimodal agent building block.
• alibaba/open-code-review: Alibaba's internal AI code review tool open-sourced, hybrid deterministic-pipeline + LLM-agent architecture with line-level comments for NPE/thread-safety/XSS/SQLi.