Coding agents & CLI

• github/copilot-sdk: GitHub ships Copilot's agent runtime (planning, tool-invoke, orchestration) as embeddable SDK, Python/TS/Go/.NET/Java/Rust, same engine as Copilot CLI.
• Claude Code v2.1.214: fixes dir/** allow-rule bug that auto-approved writes outside cwd, closes a Windows PowerShell 5.1 permission bypass, Bash checks now fail closed.
• Claude Code v2.1.212: /fork now spins conversation into a background session (own row in claude agents), old subagent launch renamed /subtask.
• Claude Code v2.1.211: new --forward-subagent-text flag streams subagent text/thinking into stream-json output, fixed permission-preview spoofing via bidirectional-override/zero-width chars.
• Claude Code v2.1.210: live elapsed-time counter on collapsed tool summaries, warns on Write(path)/NotebookEdit(path)/Glob(path) perm rules, use Edit/Read instead.
• stablyai/orca: ADE runs Codex/Claude Code/OpenCode/Pi side by side, each in own git worktree, one dashboard, cross-provider usage tracking.
• bytedance/deer-flow: open-source long-horizon superagent harness (sandboxes, memory, skills, subagents) hit #1 GitHub trending on v2 launch.
• alibaba/open-code-review: Alibaba's internal review bot open-sourced, deterministic pipeline + LLM agent, line-level comments, fine-tuned ruleset for NPE/thread-safety/XSS/SQLi.
• Port22: run Claude Code/Codex from phone.
• AgentMicro: macOS menu-bar live status for Codex tasks.
• SnapState: persistent state layer for agent workflows, thin HN post.
• diegosouzapw/OmniRoute: free MIT AI gateway, one endpoint to 290+ providers/500+ models, quota-aware fallback, claims 15-95% token savings via prompt compression.

Agent memory

• Local memory server for Claude Code that forgets on purpose: community build with deliberate TTL-style pruning instead of unbounded fact accumulation.
• TencentCloud/TencentDB-Agent-Memory: team-level memory hub, turns chats/docs/code into four shared asset types (Chat Memory, Skill, LLM-Wiki, Code-Graph), governed across agents/frameworks.

LLM releases

• Kimi K3 2.8T-A50B: Moonshot's release, largest open model yet, Opus-4.8-class capability at Sonnet-5 pricing.
• Kimi K3 on MI355X: ~952 tok/s/node, AMD beats B300 on perf-per-dollar for K3 inference.
• GLM-5.2: Nathan Lambert flags it as capability threshold, step-change for open agents.
• DeepSeek-V4-Flash-0731: frontier agent intelligence at flash-tier pricing.
• Claude Fable 5: ships with new AI-safety "fables" framing, Anthropic swapped silent research-query manipulation for a classifier like other safety domains.

LLM mechanics & economics

• Persistent State Machines: claims formal proof of INT4 in-memory attention cells, HN dismissed as buzzword salad, no substance, skip.
• Frontier post-training recipe review: Lambert + Finbarr Timbers on what an Olmo-style recipe needs to reach frontier post-training.
• Some ideas for what comes next, May 2026: open/closed balance, power struggles, AI economics turning real.
• CostPerPrompt: live pricing calculator across 232+ models, HN notes it skips token-efficiency/cache-behavior nuance.

Vibe-coding & voice

• CoreBunch/Instatic: self-hosted agentic CMS, Webflow/Framer alternative, visual editor + Bun server outputs clean static pages.
• huggingface/speech-to-speech: open local voice-agent pipeline VAD→STT→LLM→TTS, OpenAI Realtime-compatible WebSocket API, fully swappable/local-capable.