Praxist

Praxist is Sapient Intelligence’s open (Fair Source) autonomous research system for measurable, computer-executable problems. You keep the runnable project, metrics, and evaluator; Praxist owns orchestration: parallel research peers, multi-metric ranking, quality-diversity allocation, optional Deep Innovation Gate, resume/replay, and resource scheduling. On the 75-task MLE-bench suite their paper reports 49 golds at about $3k model spend versus 34 golds for a Claude Code baseline near $38k. Codex-native mode can run off a saved Codex login; Claude Code skills and CLI control are also supported.

Most agent demos chat once and forget. Praxist treats research as a multi-generation process with evaluators, evidence lanes, and a planning panel — so long campaigns stop re-learning the same dead ends and you can inspect why a claim stuck.

Install with `pip install "praxist[agents,codex]"` then `praxist setup --interactive --install-skills codex` (or the Claude Code install path). Point Codex at an already runnable research repo and run `$praxist-takeover` with a tight brief: objective, metrics, constraints, peer/generation budget. Monitor with `praxist --monitor --latest`; stop/resume by run id. Ship example tasks via `praxist examples install rocket_booster_recovery` before wiring your own harness.

Not another coding TUI or one-shot AutoML sweep. It is a lineage-centered research graph: later generations inherit validated mechanisms and failed claims, evidence stays typed and inspectable, and the product is built for multi-day R&D budgets rather than single agent sessions. Note Fair Source commercial terms once org revenue crosses $1M.

• Public demo available

• LoopX ↗
• OpenWorker ↗
• qm ↗