When Mayank and I were originally ideating the concept that became SquareDiff.com, we started with a simple thesis we kept observing while doing knowledge work: different models produce very different results for different tasks, and finding the best output often meant running many of them and comparing the differences by hand. That raised the next question. If agents are the future, shouldn't we be comparing their answers and their harnesses too? Could we even query products like Manus, which were essentially testing different harnesses themselves?
Chat.SquareDiff.com let anyone test different models, agents, and harnesses in a split-screen view that highlighted their differences. Put in a prompt, select the models and agents you want to test, and run them all at once. Agnostic to any model.
- Designing a split-screen interface and UX that keeps four or more concurrent outputs readable and comparable at a glance.
- Querying models, agents, and even applications like Manus, so entire harnesses could be tested, not just raw model calls.
- Handling staggered completion gracefully: runs finish at different times, and each pane updates independently while the overall run is still in flight.
The prototype was built with Next.js, React, and a shadcn-style component system, and deliberately simulated concurrent agent runs first to nail the generate, run, and compare interaction before wiring in live APIs. In the future, we would have liked to build out more: model preferences learned from which answer a user liked best, enabling intelligent routing of knowledge work to the model best suited for it.