sapientinc/praxist

Praxist: meet your personal R&D team

Praxist is an autonomous research system for measurable, computer-executable research. It coordinates parallel research peers, task-owned evaluation, durable evidence, and generation-to-generation synthesis.

Praxist treats research as a persistent process rather than a sequence of disconnected prompts. Use it when a project already runs and its objective is measurable, but the best path forward is still unknown.

Install Praxist

Install the complete runtime integrations and finish first-use setup with one command:

python3 -m pip install --index-url https://pypi.org/simple "praxist[agents,codex]" && praxist setup --interactive --install-skills codex

The local wizard covers the Fair Source License, User Agreement, privacy, runtime profile, masked credentials, Codex skills, writable examples, and readiness checks. It does not select a research project or launch a run. For Claude Code, use the host-specific one-line command.

For an agent-managed installation, open Codex:

codex --yolo

Then ask it to install and configure Praxist using the packaged OOBE runbook, and to stop after readiness checks.

Before starting research, read the Quickstart and Your First Task. They describe the separate takeover step and the project contract it creates.

Choose Codex-native mode to use an existing Codex subscription without an API key. For sustained research, Praxist generally favors open-source model APIs with a high observed cache-hit rate. The setup wizard also supports other API-backed profiles.

Use Praxist Through Codex

We recommend Codex as the interface for operating Praxist. Praxist is not a replacement for Codex: Codex remains the interactive agent that understands your project, communicates with you, and uses development tools. Praxist adds the persistent research loop, parallel peers, evidence protocols, scheduling, and lifecycle control.

After installation, open Codex in the root of an already runnable research project and invoke $praxist-takeover. The takeover skill inspects readiness, creates or repairs the task harness, validates its evaluator and evidence contract, and launches the run after the required gates pass. A precise brief produces a better research plan; include the objective, metrics, constraints, resources, exploration choices, and whether launch is authorized.

Other bundled skills:

See Agent Skills for invocation syntax and the generated Skills Reference for the complete contracts.

What Praxist Provides

Praxist And The Task Project

Praxist contains no task-specific scientific assumptions. A task remains the single source of truth for what should be tested and what counts as valid evidence.

Operate A Run

praxist status --json
praxist --monitor --latest
praxist stop 
praxist resume 

Ctrl-C closes only the monitor; it does not stop the research run.

Examples And Templates

praxist examples list
praxist examples install rocket_booster_recovery
praxist examples install rocket_booster_recovery_rust

Complete examples are writable reference projects. templates/tasks/ contains replaceable scaffolding for building new task harnesses. The two Rocket Booster Recovery examples demonstrate the same research problem through Python/JAX and native Rust implementations.

Requirements

Task-specific datasets and compute dependencies remain owned by the task project.

See the platform support matrix for the difference between release-qualified hosts and compatibility targets.

Documentation

Read the Praxist documentation or open it with:

praxist docs

No local documentation server is required.

Contributing To Praxist

Read the contribution guide together with the Code of Conduct before participating. Source maintainers should also follow the architecture and maintenance contract in AGENTS.md. See the Privacy Notice for Praxist’s data-handling terms.

Contact: [email protected]

FAQ

Citation

If you use Praxist in your research, please cite:

@misc{li2026praxistexperimentalartifactssolution,
      title={Praxist: From Experimental Artifacts to Solution Lineages},
      author={Jin Li and Ahmed Murtadha and Zhiyu Wang and Qiwen Chen and William Chen and Yifei Wu and Guan Wang and Andy L. Siy and Jiayi Yang and Mengsha Huang and Wenhao Li and Yixuan Liu and Shuailin Pan and Mingli Yuan and Sen Song and Yuhao Sun},
      year={2026},
      eprint={2608.25955},
      archivePrefix={arXiv},
      primaryClass={cs.MA},
      url={https://arxiv.org/abs/2608.25955},
}