Agents that operate software the way a person does.

Beyond All Reason is an open-source real-time strategy game. A player manages an economy, builds an army, and fights, all at once and in real time. That makes it a demanding testbed for an agent that has to read a situation and act on it continuously.

Our agent is a scripted executor guided by our own open-weight decision model, fine-tuned on NVIDIA DGX Spark hardware through self-play. The model reads a text summary of the game state and answers a fixed set of questions; the executor turns those answers into orders. Together they already beat the game's built-in easy AI most of the time. Everything we learn about state, timing, and evaluation carries over to agents for ordinary software.

Testbed

Beyond All Reason, an open-source RTS game

Agent

A scripted executor guided by our own open-weight decision model

Training hardware

NVIDIA DGX Spark

Result so far

The agent wins about 70% of games against the game's easy AI

Application

The same loop extends to agents for other apps and websites

Products we built and run. They pay for the research and keep us honest about what works outside a lab.

• VerbalBuilder

  Adaptive GRE, SAT and ACT verbal practice. Over 1,400 reviewed questions, AI essay grading, and spaced-repetition vocabulary with an audio podcast.

  verbalbuilder.com

• TypeVocab

  A typing speed test built on real GRE, SAT and ACT vocabulary. See each word's definition and hear it pronounced as you type.

  typevocab.com

• OpenTriangle

  Appointment scheduling for local businesses. Customers book directly, or through AI assistants such as ChatGPT and Claude.

  opentriangle.aiApp Store

Everything here, from the game agent to the products, is written with Claude Code. We keep a short loop between an idea and something running.

If you are working on agents that use software, or want to talk about any of this, write to contact@bearmountainlabs.com.