The future is autonomous.

Deep learning has produced remarkable results, but it demands enormous amounts of data, resists inspection, and does not transfer well. Trained in simulation, it tends to absorb correlations specific to that simulator rather than reusable abstractions, so behavior degrades as soon as conditions change.

We take a different route. Rather than training a network to act directly, we search for the program that does. Learned models guide that search, and candidates improve over successive generations, judged on how they actually perform rather than on how closely they fit a dataset.

Because candidates are judged across many tasks and environments rather than tuned to a single one, the search favors strategies that generalize over shortcuts that happen to work in one setting. What it produces is compact and legible: a policy you can read and check before it runs, arrived at without the vast real-world datasets that end-to-end methods require.

We are a team of physicists, mathematicians, statisticians, and engineers, backed by investors who understand this problem deeply. Argos Research is currently in stealth: our work is not yet public, and we intend to share research and release code as it matures.

Technical brief — “Programmatic AI: A Paradigm for Interpretable, Transferable Machine Intelligence” Alsibassi & Ismail, Argos Research, 2026. Available on request.