Model Releases

  • Zyphra/ZONOS2 🤖 — Zyphra’s new TTS model released under Apache‑2.0, giving you a fresh voice option for synthesis tasks.

Open Source Releases

  • AgenticBridge 0.1.0 🛠️ — SQLite‑backed MCP server that acts as a shared memory bus for AI coding agents, persisting context across sessions.
  • nvidia-nat-langchain 1.9.0a 🛠️ — Official NeMo Agent Toolkit subpackage that plugs NVIDIA’s agents into LangChain/LangGraph, easing production‑grade LLM agent builds.

Research Worth Reading

  • ITNet 📄 — Proposes ITNet, a learnable integral transform that mathematically unifies convolutions, recurrences, and transformers.
  • Diffusion Language Models 📄 — Experimental analysis of DLMs, an autoregressive alternative that generates text via iterative denoising, with benchmarks against classic LLMs.
  • Deontic Policies 📄 — Introduces a runtime governance framework for LLM‑driven agents, encoding obligations, permissions, and prohibitions beyond simple auth.
  • Hidden Anchors 📄 — Explores why multi‑agent LLM deliberation improves reasoning, identifying “hidden anchors” in early rounds that skew later consensus.
  • Uncertainty Decomposition 📄 — Extends aleatoric/epistemic uncertainty for interactive agents, providing decomposable uncertainty that lets agents ask clarifying questions proactively.
  • Post‑Training Pair Selection 📄 — Studies optimal pairing strategies for preference‑based post‑training (DPO/RLHF), giving actionable guidance on which completions to compare.

AI Dev Tools

  • LTX-2 🤖 — Official Python package for inference and LoRA fine‑tuning of Lightricks’ LTX‑2 audio‑video model, supporting joint AV generation.
  • hermes-agent 🤖 — NousResearch’s adaptive AI agent that grows with the user, integrating with Claude Code, Codex, and other harnesses.
  • Claude Code v2.1.183 🛡️ — Blocks destructive git commands and terraform destroy unless explicitly requested, plus guards against unauthorized commit amendments.
  • cc-switch 🛠️ — Cross‑platform desktop app that unifies management of Claude Code, Codex, OpenCode, OpenClaw, Gemini CLI, and Hermes Agent, built with Rust/Tauri and WSL support.
  • nanobot 🛠️ — HKUDS’ lightweight open‑source AI agent for tools, chats, and workflows, supporting Claude Code, Codex, and other frameworks with simplicity and extensibility.

Today’s Synthesis

The agent ecosystem is quietly standardizing around shared memory and unified control. AgenticBridge gives you a SQLite-backed MCP server that persists context across sessions — plug hermes-agent and nanobot into it, and you’ve got multiple specialized agents (one for refactoring, one for tests, one for docs) that actually remember what the others did. cc-switch then becomes your mission control, letting you hot-swap between Claude Code, Codex, and these open agents without losing thread continuity. The missing piece? Governance. Deontic Policies shows how to encode obligations and prohibitions at runtime — pair that with Claude Code’s new guardrails (blocking terraform destroy and unauthorized commit amendments) and you can enforce “no force-push to main” at the policy layer, not just the config layer. Start small: spin up AgenticBridge locally, register two agents with distinct roles, and write one deontic rule — “agents may not delete files without explicit user confirmation.” You’ll have a governed, memory-sharing multi-agent loop running in an afternoon.