Manages an Obsidian vault through a crew of eight AI agents. You speak naturally and the system automatically files, connects, and triages your information.

https://github.com/gnekt/My-Brain-Is-Full-Crew https://x.com/tom_doerr/status/2084058725506073052

Give your AI assistant a persistent memory to capture, search, and reuse project knowledge across agents and repos.

https://github.com/0xranx/OpenContext https://x.com/tom_doerr/status/2084089693365325995

Use realistic https://app.diagrams.net/ shapes for network gear from APC, MikroTik, Ubiquiti, Sophos, and Supermicro, standardized to consistent U scaling.

https://github.com/Fabl0s/Draw.io-Shapes-Collection https://x.com/tom_doerr/status/2084104082428473355

Converts web URLs, PDFs, and media files into clean Markdown, automatically stripping ads and navigation while supporting dynamic content rendering.

https://github.com/AeternaLabsHQ/pullmd https://x.com/tom_doerr/status/2084164751098089889

PentAGI automates penetration testing by combining AI-driven planning with a fully isolated Docker sandbox.

- Runs fully autonomous pentests with optional execution monitoring
- Bundles 20+ professional tools: nmap, metasploit, sqlmap
- Long-term memory stores research results and successful approaches
- Supports multiple LLM providers: Ollama, OpenAI, Anthropic, Gemini, Bedrock

Big monolith is the best architecture. https://x.com/systemdesignone/status/2084290853523108244

NVIDIA is bringing OOP to AI agents!

(bookmark this)

They open-sourced a new framework called NOOA, built on the idea that an agent can be defined as a single Python class instead of being spread across prompt files, tool schemas, callbacks and a workflow graph.

Methods are the actions the model can take, fields hold its state, docstrings become the prompts, and the return type is the contract its answer has to satisfy.

The clever bit is what an empty method means here.

If the body is just three dots, an LLM fills it in at runtime. If it has real code in it, it runs as ordinary Python with no model involved.

For example:

@𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆(𝗖𝗼𝗱𝗲𝗔𝗰𝘁𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝘆())
𝗮𝘀𝘆𝗻𝗰 𝗱𝗲𝗳 𝘁𝗿𝗶𝗮𝗴𝗲(𝘀𝗲𝗹𝗳, 𝗺𝘀𝗴: 𝘀𝘁𝗿, 𝗼𝗿𝗱𝗲𝗿: 𝗢𝗿𝗱𝗲𝗿) -> 𝗧𝗶𝗰𝗸𝗲𝘁:
"""𝗧𝗿𝗶𝗮𝗴𝗲 𝗰𝘂𝘀𝘁𝗼𝗺𝗲𝗿 𝗺𝗲𝘀𝘀𝗮𝗴𝗲 𝗻 𝗰𝗿𝗲𝗮𝘁𝗲 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 𝘁𝗶𝗰𝗸𝗲𝘁."""
...

So one class shows exactly where the model is allowed to make a judgment call and where the behaviour is fixed, and that line is one method wide.

The second idea is the one I found more interesting, and it's about how data gets in.

Say a method is called with a list of 100 records. The model never sees the list itself, only the variable name, its type, its length, and a few items from each end.

The actual list stays live in the session the model writes code against, so it can loop over all 100 while only ten of them ever reach the prompt.

That quietly changes what an agent can handle. The limit becomes the machine running the code rather than the context window, so a method can take a table with millions of rows and the prompt still carries a short preview.

Most agent setups solve this by dumping data into files, which works, but then the model has to read plain text and rebuild the structure in its head every time.

The payoff shows up in cost. On SWE-bench Verified with GPT-5.5 it scores 82.2% against 78.6% for OpenCode, using roughly 28 model calls and 1.1M tokens per task where a comparable agent needed 66 calls and 2.2M.

None of these models were trained on this interface, and they still passed 97.9% of 4,400 tests on it, mostly because it is just Python they already know well.

NOOA is inspired by PyTorch, which showed that a powerful runtime can still present users with a simple Python programming model.

Paper: https://arxiv.org/abs/2607.20709
Repo: https://github.com/nvidia-nemo/labs-OO-Agents

1 employee. 1 week. $10K in model tokens.

Built the GenOffice Alpha.

It’s now open source and being built in public. Test it, report bugs, help shape the roadmap, and earn 1,000+ credits.

GitHub:
https://github.com/genspark-ai/genoffice

Download:
https://www.genspark.ai/genoffice

Join the Alpha:
https://www.genspark.ai/genteam/invite-link/invite_link_069269651ea1457091451e909a88f240?token=QsRgYXmXQ3pXDu9xvZ-C-gFjhJI3S-LNRSH5wPJIpyI