Wild looking back at this now. This was an early glimpse of what eventually became a much larger part of Ares. Since then I’ve been building out deterministic model routing, semantic memory injection, governed tool execution, receipts/evidence, context management, heterogeneous local compute (I combined my crate with NVIDIA PAIR -- how I came across the bug that i patched and that merged into the NVIDIA/PAIR repo), and resumable agent execution. The original goal hasn’t changed: make capable agent systems more efficient, inspectable, and increasingly local-first. What started as “how much can I run in 16-32GB?” has turned into “how much of the agent stack can I make deterministic, verifiable, and owned by the operator?” -- I owe everything about my drive and direction right now to the low key encouragement by @Teknium