Multi-agent handoff notes: when Claude reads memory as orders
In stillwet's painter chains, Claude read earlier notes as instructions and missing notes failed silently, so label, log and verify agent handoffs.
ldraw-nova: AI agents build LEGO models by writing Python geometry
In ldraw-nova, models write a plan and a Python generator instead of raw coordinates, but nothing checks physics or stability.
Huseyin Kaplan · 1 min read
An AI agent can now design a LEGO crane with over a thousand parts. Nothing it runs can tell you if the crane would stand.
ldraw-nova, a Show HN project this week, has GPT-6 Astra or Claude Opus 5.5 build models in LDraw, a text format where each line places one part with a position and a rotation matrix.
Early attempts asked models to write those numbers directly. Close, never quite correct.
What worked: the agent writes a plan file, then a Python generator, and the script outputs the parts. Models are weak at geometry and strong at code that does geometry.
Its tools catch collisions and gaps and render images the agent inspects. Nothing tests physics or stability.
For any agent that outputs geometry:
- Plan in a file you can diff
- Code computes the numbers
- Render, inspect, fix
- Write down what you don't check
Let the model write the math, not do it.
