Does dynamic loading
actually help?

We built MOAH to test a simple question: can an agent keep access to more capabilities without paying the context cost of exposing everything at once?

57 runs. Same model. Same tasks.

MetricPiMOAHOpenCode
Fresh tokens1.00×1.44×3.64×
Processed1.00×1.30×6.96×
Success17/1917/1914/19
Median E2E4.9s6.1s11.2s

MOAH isn’t smaller than Pi. That’s not the point.

Pi starts with a deliberately minimal tool surface. MOAH explores whether a larger capability library can remain available while keeping the active working set small.

Against OpenCode

~60%

fewer fresh tokens

Median fresh input tokens across the benchmark suite.

~81%

fewer processed tokens

Median processed tokens across the benchmark suite.