Praxist

Praxist is Sapient Intelligence's autonomous research runtime for projects that already run and already have a metric. Docs: praxist.sapient.inc/en/docs. The repo sapientinc/PRAXIST is Fair Source License Agreement 1.0, created 2026-08-27, Python. On 2026-09-10 GitHub listed 6,280 stars and 602 forks. Paper: arXiv:2608.25955. README: meet your personal R&D team.

Compare DeerFlow if you wanted a general multi-agent research app, LangGraph if you wanted a graph runtime you wire yourself, or DeepSeek Harness if you wanted a coding-agent harness rather than a multi-generation lab.

Key Features

• Parallel research peers: competing hypotheses in the same generation, then synthesis into the next agenda.
• Durable evidence lanes: incubator, frontier, and Gems. Invalid or suspicious evaluator output is dropped.
• Task owns science: Praxist owns orchestration, scheduling, replay, and plugins. The project owns the objective, code, evaluator, metrics, and baselines.
• Bundled skills: $praxist-takeover, praxist-control, praxist-diagnostic, praxist-scientific-research, plus Codex-native no-key takeover. Claude Code has a host-specific one-liner in the install doc.
• Stops when the project is not ready: README says it will not silently download datasets, invent a simulator, or fabricate a baseline.

Limitation: Fair Source is source-available, not OSI. Orgs at or above US$1M aggregate annual revenue must contact Sapient for a commercial license. Higher-ed and nonprofit research are exempt. 6,280 stars in about two weeks is a heat signal, not an audit. You need CPython 3.11+ and a project that already evaluates.

Use Cases

• Labs with a runnable baseline who want parallel search plus provenance, not another chat thread.
• Codex or Claude Code users who will keep the interactive agent and let Praxist own the overnight loop.
• People who should use LangGraph instead if they need to author the graph and do not have a measurable evaluator yet.

Pricing

External publications of generated outputs must keep the attribution "Praxist by Sapient Intelligence". Internal use does not need that line.

Getting Started

1. python3 -m pip install --index-url https://pypi.org/simple "praxist[agents,codex]" && praxist setup --interactive --install-skills codex
2. Open Codex in an already runnable research project and invoke $praxist-takeover with the objective, metrics, peer count, and whether launch is authorized.
3. Monitor with praxist --monitor --latest. Ctrl-C closes the monitor only.
4. praxist examples install rocket_booster_recovery if you need a writable reference task.

First-party resource: installation and praxist doctor.

Frequently Asked Questions

Is Praxist MIT?

No. README FAQ: Fair Source License Agreement 1.0. Source is public. Commercial use below US$1M revenue is free; above that, negotiate.

Will it invent my dataset?

No. If the baseline or data is missing, it stops and tells you.

Does it replace Codex?

No. README: Codex stays the interactive agent. Praxist adds the persistent research loop.

Alternatives

• DeerFlow: open multi-agent research product, not a generation scheduler on your evaluator.
• LangGraph: you define the graph.
• CrewAI: role-based crews without Praxist's evidence lanes.
• DeepSeek Harness: coding-agent harness, not multi-generation R&D.

Tips

1. Quote Fair Source, 6,280 stars, CPython 3.11+, and the US$1M threshold from pages checked 2026-09-10.
2. Write the takeover brief with metric direction and constraints. Vague briefs make weak harnesses.
3. Run praxist doctor on macOS. Linux 3.11/3.12 is the continuously release-tested matrix.

Conclusion

Praxist is the Fair Source loop for research that already runs: parallel peers, evidence, and Codex or Claude Code skills. Start at the docs. If you still need to invent the evaluator, open LangGraph first.