Drawgent puts a coding agent on a live Excalidraw canvas
Labs stayed quiet before DevDay, but builders still got open 7B robot weights that top the leaderboard — plus fresh tools to design and stress-test agents.

Copy markdown
A coding agent that lives on your Excalidraw board
Drawgent, trending on Hacker News, wires a coding agent into a live Excalidraw canvas over MCP: you sketch rough architecture or UI, and the agent arranges, refines, and writes code against the diagram instead of a scrolling chat log. It's open source on Tangled. One caveat from the thread — Excalidraw ships its own first-party MCP, so compare the two before you commit.
Open 7B robot weights that top the leaderboard
Black Forest Labs — the team behind the FLUX image models — released FLUX 3 Action, a 7B open-weights "world action model" for robots. It leads the RoboLab-120 benchmark at 42.92% versus the prior open best of 36.82% (+6.1 points) while using 44% of the parameters and running 1.43x faster, and it beats π0.5 on real-time factor. Weights, fine-tuning recipes, and a LeRobot/SO-101 example are out — final license terms and full training code are still pending. If you touch embodied AI, you can fine-tune it on your own robot demos for pick-and-place, sorting, or assembly. Dropped Sept 23.
A browser arena to battle-test your agents
TinyAIArena, also on the Hacker News front page, pits LLM agents against each other in code-generation matches you can watch and replay in the browser — a fast, visual gut-check on which model actually wins your kind of task before you wire it in. Fair warning from commenters: it reads more as a vibe check than a rigorous benchmark.