The defining challenge of our generation is directing autonomous intelligence.
As AI becomes more capable than us, the challenge shifts from building intelligence to directing it. But every model behaves differently and the way we get them to do what we want changes across models, versions, and even reasoning settings.
Propensity Labs is building the compiler between human intent and superintelligence.
Intelligence is advancing faster than our ability to direct it.
Every model has its own quirks and failure modes. One games the test to make it pass. Another loops on the same command with high verbosity, burning through tokens. Instructions that work on one model may fail on another and even changing a version or reasoning level can change the result.
Today we aim to make agent behavior independent of the underlying intelligence
We systematically study how frontier and open models behave, how that behavior changes across settings, and which interventions work. Every model we study adds to our understanding of how intelligence needs to be steered.
We are applying that knowledge first to model migrations: adapting existing agents to new models without rebuilding them from scratch.
Tomorrow we become the compiler between human intent and machine intelligence.
Compilers separated software from the hardware underneath. We believe AI needs the same abstraction.
Humans should be able to specify what they want done without understanding every model, prompt, tool, reasoning mode, or coordination strategy required to make it happen. The compiler handles how the intelligence underneath executes that intent.
As agents grow into fleets and eventually perform most economic work, the most important problem for humans to focus on is ensuring they're doing what we want them to. The interface between human intent and machine intelligence will become the most important piece of software in the world. We're building that.