Why it matters
Directly on the open-versus-closed model tradeoff and agent reliability engineering — the two questions that decide most real inference bills.
The tokenmaxxing angle
One talk covers evals putting open models near closed frontier models on tool-use and instruction-following at lower cost and latency, plus how to swap them in — the core routing decision tokenmaxxing tracks.
From the organizers
Sydney Runkle presents a four-loop agent harness design; Mason Daugherty compares GLM 5.2 and MiniMax M3 against closed models on LangChain's own evals. Harvard Square, 6pm.