Hey Devs! :) This week is all about making AI agents more practical: more portable, more efficient, and better connected to real tools.
GitAgent
GitAgent is an open standard for defining AI agents in a portable, git-native way. It was launched by and positions itself as a way to reduce fragmentation across ecosystems such as LangChain, CrewAI, Microsoft AI Foundry, Snowflake Cortex, Databricks Agent Bricks, n8n, and Zapier. This could make agents easier to version, move, and review across different stacks.
NVIDIA Nemotron-Cascade 2
NVIDIA released Nemotron-Cascade 2, an open 30B MoE model with 3B active parameters. The big story here is efficiency: strong reasoning and agentic performance without needing a huge always-active model. This is another sign that smarter, smaller models are becoming more serious.
Read more: https://research.nvidia.com/labs/nemotron/files/Nemotron-Cascade-2.pdf?ref=arwriter.ai
Google Colab MCP Server
Google Colab now offers an open-source MCP server, making it easier for local AI agents to connect to Colab runtimes and GPUs. This is a practical step for developers who want to combine local agent workflows with cloud computing.
Claude Code Telegram Plugin
Anthropic’s official Claude plugins repo now includes a Telegram plugin. It lets Claude Code connect to Telegram via MCP, opening up interesting automation and remote-assistant use cases. Just make sure access settings are configured carefully.
Read more: https://code.claude.com/docs/en/channels
GitHub Dev Day Perth
On a personal note, I was a speaker at GitHub Dev Day in Perth. My talk was about orchestrating multiple agents in VS Code. It was a great event with a lot of good content and many new connections. If you want more details about the talk or the material I shared, feel free to reach out.