PRAXIST launches generational autonomous R&D

PRAXIST’s most interesting idea is treating experiments as cumulative research evidence rather than disposable agent sessions. The approach could make long-running AI research more auditable and less prone to repeating dead ends, though its benchmark claims remain company-reported.

• –Agents explore competing hypotheses and implementations in parallel
• –Typed evidence graphs preserve mechanisms, failures, provenance, and unresolved questions
• –Task-owned evaluators keep domain metrics and scientific constraints outside the orchestration layer
• –Codex and Claude Code integrations make the system accessible through familiar coding-agent workflows
• –The source-available Fair Source license may limit adoption compared with truly open-source research frameworks