I took Vercel's dogfood skill and turned it into a general bugbash for all codebases. Now it can test CLIs, TUIs, Backends, Services and other programs without a Web UI. Tonight I tried to pair it with timeboxed iteration and nested subagents setup, it worked quite well. The

Pi + custom harness is a way to go. Your harness should be a DAG that is assembled for each specific task automatically.

I think we're in for a big surprise.

Made some updates for the debug UI for our internal agent testing tool. 1. A chat UI that allows you to specify "extra" payload for chat/response or message completions (yes, all three major APIs) is extremely convenient 2. Debug UI should have "copy", "resend", and edit

Gemma 4 E2B is an extremely interesting model, I dare to say even more so than the rest of the bunch. It's similar to when Google released Gemma 2 2B and everyone said "no way this thing is coherent", but it was. E2B can write you software in realtime to run on your phone,

Google DeepMind Such a great release. Each of the models is interesting in its own way. 31B - packing what was recently frontier capability (~Sonnet 4 level) into such a small size, great response to the Qwen 3.5 27B 26B - great for actual use on most of consumer devices, to power up all the

New Gemma 4 models cutoff date is January 2025. Remember what the world was like back then? That's this model perception of reality and will always be (unless they release new checkpoints).

You downloaded a new amazing model and run a few prompts, what's next? 1. Add a Web RAG, to give model access to current data, SearxNG is a perfect self-hosted service for that. 2. Add it an ability to work with files and run commands. Do not do it on your own machine, it's not

Mintlify gave its agent access to docs via just-bash + custom filesystem interface on top of Chroma DB.

Zeneca🔮 I really tried to make OpenClaw work with Kimi 2.5, but it was unusable with anything smaller than Sonnet 4.6... Hermes, Qwen 3.5 35B drives is mostly without issues. So yeah, a pretty big difference.

Harveen Singh Chadha Nobody noticed because those MAI models are for transcription and image generation and only available via API, so ofc Gemma 4 took the scene with Apache 2.0 :)

David Ondrej How about this: ``` harbor pull gemma4 harbor up searxng openterminal harbor open ``` Now you have Open WebUI + Web RAG + persistent sandbox for the LLM to use to work with files, all containerized without risking your host if a clanker goes wild :)

Yesterday, after Gemma 4 release I wanted to test it, but immediately spotted a few issues with harbor trying to run this new model, firstly there was no way to run actual latest llama.cpp as we're relying on pre-built Docker images, secondly the was no way to specify CUDA arch

I suggest we rename solopreneur to slopopreneur at this stage.

Like many of you here, I'm running my lifeos with agents now, I also have a dashboard aggregating all kinds of useful stats that I want to track, to conveniently access them in one place. Tracking some software releases is one of such tasks, so a couple of weeks ago I added

Don't use AI if you can't take a yes for an answer.

DROP EVERYTHING > install Harbor > harbor pull unsloth/gemma-4-31B-it-GGUF:Q4_K_M > harbor up llamacpp searxng webui > open Open WebUI > load Gemma 4 Now your local model has a UI, web search, and a sandboxed stack

Just updated `run-llms` skill, you can now manage Harbor with your agent (I recommend Hermes / OpenCode + Qwen 3.5 35B or Gemma 4 26B and above) "What models do I have" "Pull unsloth quant for Qwen 3.5 9b" "Switch to optimised llama.cpp image for Strix Halo" "Start hermes