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.
Pi 1.0 agent harness: the Claude Opus planner drove 95% of the cost
In Pi 1.0's demo, Opus used a fifth of the tokens yet cost 18x more than GPT 6 Luna, so log cost per model and count cache misses.
Huseyin Kaplan · 1 min read
In Pi 1.0's launch demo, Claude Opus handled about a fifth of the tokens and about 95% of the bill.
Earendil shipped Pi 1.0 yesterday, the open source agent harness at the top of Hacker News this week.
The demo builds a virtual model. Opus plans, GPT 6 Luna implements, and Jev, a non-LLM model, decides when to hand over.
Pi's /session command splits cost per model. Opus used around 20K tokens. Luna used around 76K. Opus still cost about 18 times more.
The cheap model did the long typing. But the planner's share sets the ceiling on what you save.
The handover has a price too. The same screen shows one cache miss, a few thousand tokens billed again.
Before you split planner and implementer:
- Log cost per model, not per session
- Count cache misses at every switch
- Cap planner output per task
- Check the implementer doesn't plan again
The cheap model types. The expensive one writes the bill.
