The open-source workflow has a human approve a plan, one agent build it in rounds, and a reviewer agent from a different vendor sign off, with deterministic hooks routing the work. Its authors' 118 public postmortems show review catches unsupported claims, but knowing when to stop is still the weak spot.
Vega pairs a frozen Qwen3.5 0.8B reader with a small trained engine to return choices, 0–1 scores, or true/false probabilities with an explicit abstain flag, plus 73k token context and image input. It's Apache 2.0 with a 4B variant, installable via pip install vegaml…
Drex 1.5 takes a state plus typed questions (choice, yes/no, score) and returns a probability for every option in one pass, with up to 131k tokens of context. Weights are on Hugging Face with llama.cpp and Ollama support, and Nace says it tops the Decision Index among models under 10B…
Redis Docs MCP at redis.io/mcp needs no API key and exposes three tools: search (up to 8 pages), fetch (full page), and ask (a synthesized answer with sources). It's read only and separate from mcp redis , which connects agents to your own database…
Scheduled agent tasks now live under Automations, where you can fire one immediately with Run Now or pick Create with Prompt to have the agent help set up a new one. The release also adds ⌘/Ctrl+K search across conversation contents and folds an agent's file edits, reads, and terminal commands into one collapsible row…
The Apache 2.0 server speaks the OpenAI, Anthropic and Ollama APIs and runs GGUF via llama.cpp or MLX models. superfluid launch claude model <gguf points Claude Code at a model running on your own machine…
Spotted something we missed? Start a thread.