The new ChatGPT web app version adds a hidden option to "Use cloud browser" when enabling Agent mode

However, interestingly, this option is enabled only if the user agent matches
"ChatGPT.+Macintosh;.+ Chrome" (likely the new browser from OpenAI)

- hinting at the possibility that ChatGPT in Agent mode might be able to control either your browser or the cloud browser

A new open-source coding agent toolkit
> just solved the subscription fatigue plaguing AI-assisted development by turning free LLMs into semantic code assistants that rival Cursor and Windsurf.

Serena emerged from a simple frustration: why pay for multiple IDE subscriptions that force you to buy tokens on top of chat costs, or burn through API credits with tools like Cline and Aider? Serena leverages the Language Server Protocol to provide symbol-level code understanding without the recurring fees.

Package highlights:

(1) Symbol-level intelligence - instead of basic text search, Serena uses LSP to navigate code like an experienced developer, finding and editing the exact functions and classes you need.

(2) Complete development autonomy - agents can analyze, implement, test, and commit code changes end-to-end without constant human intervention.

(3) Framework flexibility - works as an MCP server with Claude Desktop, integrates with IDEs through extensions, or runs standalone with any LLM via Agno.

Serena works with any LLM: for developers tired of subscription fatigue or who prefer local execution due to sensitive code.

GitHub repo:
https://github.com/oraios/serena

< ai hurricane >

Open-source document OCR
has just become faster, lighter, and easier to use, without sacrificing much accuracy.

RolmOCR is a new drop-in alternative to olmOCR from the AI2, designed for developers who want better performance when parsing PDFs, scans, and complex document layouts using vision-language models.

The team behind RolmOCR made three key changes:

(1) New foundation model – upgraded to Qwen2.5-VL-7B for better raw capability

(2) No metadata – skips PDF metadata extraction to reduce prompt size, speed up inference, and cut VRAM usage

(3) Rotated training data – 15% of samples were rotated to improve robustness on off-angle or scanned inputs

In testing, RolmOCR matched or outperformed olmOCR on a wide range of examples:
* Handwritten notes - Better character recognition and reading order
* Low-contrast envelopes - Stronger generalization when metadata is unavailable
* Academic PDFs - Occasionally underperforms on tables with embedded metadata

If you’re building multimodal apps or working with messy document inputs, RolmOCR gives you a streamlined starting point.

Model (Apache 2.0):

https://huggingface.co/reducto/RolmOCR

< ai hurricane >

Python library for building "AI agents that actually follow instructions" 👀.

Parlant is an open-source AI agent framework designed to ensure LLMs reliably follow instructions.

Instead of relying on long system prompts, developers define guidelines in natural language with "guaranteed compliance" (according to their docs).

It supports tool integration, conversational journeys, guardrails against hallucination, and conversation analytics.

Agents can be deployed quickly with built-in React widgets and enterprise-ready features like explainability and compliance.

Use cases can be financial services, healthcare, e-commerce, and legal tech where consistent, rule-based AI behaviour is critical.

IMO, while instruction following is not something anyone can guarantee (yet) for LLM-based agents, the library is quite useful for building more constrained AI workflows.

Link to Github Repo:
https://github.com/emcie-co/parlant

< ai hurricane>

VibeTunnel just solved the terminal access problem
> that's been keeping developers chained to their desks while AI agents run for hours.

The new open-source framework turns any browser into a fully-functional terminal interface, letting you monitor and control Claude Code, ChatGPT, or any terminal-based AI agent from anywhere - your phone, tablet, or any web browser. No SSH keys, no VPN configuration, no port forwarding nightmares.

Highlights:
(1) Zero-friction setup - install via Homebrew or npm, click the menu bar icon, and you're instantly connected at localhost:4020

(2) Real-time AI agent monitoring - dynamic terminal titles show exactly what Claude is doing (✻ Crafting 45s, ↑2.1k tokens) with automatic activity detection

(3) Universal compatibility - native macOS app for Apple Silicon, npm package for Linux/Docker/CI environments, with full feature parity

(4) Secure remote access - built-in support for Tailscale's encrypted tunnels or ngrok's public URLs, with multiple authentication modes including SSH keys

The architecture is surprisingly elegant: a lightweight Swift menu bar app manages a high-performance Node.js server that streams terminal sessions via WebSockets to a responsive web frontend.

GitHub repo:
https://github.com/amantus-ai/vibetunnel

< ai hurricane >

Google releases the complete guide to get the best out of Nano Banana

Comprehensive developer tutorial for Nano Banana (Gemini 2.5 Flash Image), the powerful new model for image generation and editing.

📌 Inside the guide you’ll find:
🔹 How to generate images from text prompts
🔹 Techniques for photo restoration and colorization of old photos
🔹 Advanced editing with multiple inputs and conversational sessions
🔹 Best practices and prompting tips for sharper results

👉 But that’s not all: the guide also showcases real community examples and provides official resources to integrate Nano Banana directly into your projects.

https://dev.to/googleai/how-to-build-with-nano-banana-complete-developer-tutorial-646

🤖 💻 LangCode CLI

A powerful dev tool that unifies AI models for smart coding. Built with LangChain, it offers automated tasks and safe code changes through intelligent model routing and diff previews.

Check it out! 🚀
https://github.com/zamalali/langchain-code

< ai hurricane >

Just discovered TOON (Token-Oriented Object Notation), which provides an alternate token size optimized notation for use in the context of LLMs.

It offers "typically 30–60% fewer tokens than JSON." Worth looking into!

🔗 TOON repo: https://github.com/toon-format/toon

< ai hurricane >

NEW RELEASE - huggingface/skills is a universal implementation of agent context for AI tasks like training models, building datasets, and generating datasets.

- compatible with all major coding agent tools: Codex, Cursor, Claude Code, Gemini CLI.
- has integrated local script and remote long running cloud jobs.
- swaps in with skills, AGENTS.md, extensions. whatever you want to call it.
- add relevant MCP servers like the Hugging Face MCP

In practice, skills wraps around a skill based on Claude Code’s standard with instructions, resources, scripts, and examples. But it exposes entry points to all other utilities as AGENTS.md or gemini extensions.

Link: https://github.com/huggingface/skills

< ai hurricane >

To celebrate the holiday season, ElevenLabs launching 25 Days of Agents - an advent calendar of exclusive deals from top AI companies to help you build your own agents.

Every day until December 25th, we’ll unlock a new deal from partners including Railway, Cloudflare, Convex, Vercel, n8n, Hugging Face, Lovable and more.

Get access to the first two deals using the link:

https://www.25daysofagents.com/

#Release
🚀 Big news for developers and product teams! GPT-5.3-Codex is now generally available in GitHub Copilot! 🎉

GitHub just rolled out GPT-5.3-Codex — the next generation of AI coding power — right inside Copilot across VS Code, github.com, mobile, CLI and agent workflows.

👇 Why this matters
Imagine an AI that:
✨ Codes faster — up to ~25% quicker on real engineering workflows than the last version
✨ Thinks deeper — stronger reasoning and execution on complex and long-running coding tasks
✨ Fits your workflow — available in chat, ask, edit or agent modes in Copilot
✨ Enterprise ready — available for Pro, Pro+, Business & Enterprise plans (admins can enable it in settings)

Developers are already comparing its autonomy and balance to other cutting-edge agents (even some say it feels more intuitive and efficient during agentic workflows).

This isn’t just an upgrade — it’s a shift toward AI that understands your project context and helps get work done.

🛠️ If you’re building tools, architecting systems, or shipping features, GPT-5.3-Codex in Copilot is a game changer.

What will you build with it? 💡

https://github.blog/changelog/2026-02-09-gpt-5-3-codex-is-now-generally-available-for-github-copilot/