What's New
Autoresearch — the central orchestration skill that enables AI agents to autonomously conduct end-to-end research projects.
Autoresearch Skill
- Two-loop architecture: inner loop (rapid experiments with measurable outcomes) + outer loop (synthesis, reflection, direction decisions)
- Routes to all 85 domain skills automatically — agents don't need to know which skill to use
- Continuous autonomous operation: mandatory
/loop 10m(Claude Code) or cron job (OpenClaw) - Research presentations: generates HTML/PDF progress reports with optimization trajectory plots
- Project memory:
findings.mdwith "Lessons and Constraints" persists knowledge across sessions - Structured workspace: research-state.yaml, findings.md, research-log.md, literature/, experiments/, src/, data/, to_human/
- Works with any compute: local GPU, cluster job submission, cloud, or CPU-only
Demo Papers (produced by autoresearch)
- Norm Heterogeneity → LoRA Brittleness: Agent pivoted from null result on ETF overlaps to discovering norm CV as causal predictor (r=-0.99)
- RL Algorithm Brain Scan: Agent found DPO is a rank-1 perturbation, orchestrating GRPO/TRL + SAELens/TransformerLens + paper writing
Other Improvements
- WELCOME.md: cold-start bootstrap — one URL to go from zero to autonomous research
- Agent-friendly install path:
Read welcome.md and follow instructionsalongsidenpxcommand - Windows support: symlink fallback to copy when Developer Mode unavailable
- Research Ideation (21-research-ideation) added to npm installer categories
- 86 total skills across 22 categories — complete research lifecycle coverage
Install / Update
# For humans — interactive installer
npx @orchestra-research/ai-research-skills
# For AI agents — point to welcome doc
Read https://www.orchestra-research.com/ai-research-skills/welcome.md
# Install everything in one command
npx @orchestra-research/ai-research-skills install --all