Ben Swerdlow is cofounder and CEO of Freestyle, which builds computing infrastructure for AI agents. His independent work includes Brood War Bench, an evaluation that makes models act inside a real-time strategy game, and a guide to keeping agent-written code understandable.[2][3]
Start with Brood War Bench →
X · benswerd Career
Swerdlow’s founder account describes leading frontend work on the World Health Organization’s COVID-19 app effort while in high school, attending the University of Chicago, and leaving college for consulting work at Apple before founding Freestyle. The company joined Y Combinator’s Summer 2024 batch. Its current infrastructure provides virtual machines that can preserve memory and disk state across hibernation and snapshots.[2][3]
Contributions
Brood War Bench places models in StarCraft: Brood War and publishes game outcomes, timing and replay evidence. The real-time setting tests whether a model keeps responding while the world changes. A model can spend so long reasoning that it fails to build an economy or react to an attack. Swerdlow reports that none of the tested systems played beyond beginner level; winning against the other models is therefore different from demonstrating strong human-level play.[4]
Ideas
Swerdlow’s AI Code guide distinguishes small functions with explicit inputs and outputs from orchestration functions that coordinate services and side effects. It also advocates data types that reject invalid states early. The aim is to make agent changes easier to inspect and reuse by expressing domain constraints in the code rather than relying on a growing conversation to remember them.[1]
Tools, writing & talks
Writing, research & talks
Start here
Brood War Benchbw.swerdlow.devA report and replay-based comparison of models playing a real-time strategy game.
Start with a replay and compare the timing of observations and actions. No model-training background is needed; knowing the basic economy and unit controls of StarCraft helps interpret a result beyond the win column.
Author and benchmark creator[4]Original source