Catalyst

Catalyst is an open-source, self-hosted tool from Imbue for semi-autonomous scientific research, specifically in ML and deep learning theory development. Its two core modes are phenomenon explanation — where it autonomously builds a theory from observed data, or refines and corrects a researcher-supplied draft — and verifiable goal solving, where it runs an optimization loop against a user-provided verification script. The agent executes code, evaluates candidates, and iterates without requiring sign-off at each step, though ‘semi-autonomous’ in the project description signals that researchers stay in the loop for framing the problem and reviewing outputs.

The differentiating mechanism is the Darwinian evolver submodule: candidate theories or configurations compete in an evolutionary loop, with the verification script acting as the fitness function. This is not prompt-chaining dressed up as research — the agent generates, tests, and culls solutions programmatically, which means iteration speed scales with compute rather than researcher hours. The vendor documents this approach in two blog posts linked from the repository covering autonomous theory discovery and AI model research via evolution.

Catalyst fits teams doing computational research where the success criterion is measurable and scriptable — hyperparameter search with a defined metric, theory fitting against a dataset, or configuration refinement under a test suite. It breaks down when the evaluation step requires human judgment, domain knowledge that cannot be encoded in a verification script, or real-time interaction with external APIs it was not built to reach. There is no hosted API, so teams without the infrastructure to run and maintain a self-hosted agent environment carry that cost entirely themselves. Teams whose research workflows depend on GUI-driven experimentation or whose institutions restrict AGPL-licensed code in production pipelines will hit blockers before the tool’s research logic becomes relevant.