sapientinc/PRAXIST

What it solves

Praxist is designed for research projects where the objective is measurable and the code is already runnable, but the optimal path to improvement is unknown. It replaces manual iteration and simple parameter tuning (AutoML) with an autonomous, evidence-driven research loop that explores multiple hypotheses and strategies concurrently to find the best-performing solution.

How it works

Praxist transforms a project into a continuous research run by coordinating several components:

• Parallel Research Peers: Multiple agents explore competing implementations and hypotheses simultaneously.
• Multi-generation Synthesis: A planning panel analyzes evidence from one generation to shape the research agenda for the next.
• Evaluation & Evidence: A task-defined evaluator converts results into structured, durable evidence lanes (incubator, frontier, and Gems).
• Diversity Mechanisms: It uses Quality-Diversity (QD) and a Deep Innovation Gate (DIG) to prevent the system from getting stuck in local optima.
• Orchestration: It manages resource scheduling, lifecycle control (start/stop/resume), and end-to-end provenance to ensure results are reproducible.

Who it’s for

It is built for researchers and developers who have a baseline runnable project with a clear metric for success and want to automate the iterative process of discovery and optimization.

Highlights

• Autonomous Iteration: Manages the full research loop from hypothesis generation to synthesis.
• Coded Evidence: Every improvement is backed by a lineage of evidence and a verifiable evaluation path.
• Integration with Codex: Recommended to be operated via Codex for interactive project understanding and tool use.
• Language Agnostic: Supports projects in various languages (e.g., Python/JAX, Rust) as long as they are executable and measurable.