autoresearch

# Autoresearch v2.0. The Self-Improving Skill Loop > **Lineage**: Karpathy autoresearch (v1) → Stanford Meta-Harness (Lee et al., 2026) > **Core insight**: Scores-only hill climbing plateaus. Full execution trace access + counterfactual diagnosis converges faster and finds better optima. ## Source & Pairs with - **Primary lineage refs**: Karpathy autoresearch loop pattern (2025) + Stanford Meta-Harness paper (Lee et al., 2026) - `skills/loop-experiment/SKILL.md`. sister autoresearch pattern lift (from a public autoresearch sandbox pilot) - A trio audit of public autoresearch repos (gepa / uditgoenka / zkarimi22). - A content-curator pilot's round-1 results (Critic+Generate-B wins) - `scripts/eval/skill_failure_tracker.py`. the failure ledger that feeds autoresearch's regression diagnosis - `memory/Decisions.md`. past autoresearch ship decisions for calibration - Sister skills: `skills/critic/SKILL.md` (the validated subagent that pairs well with autoresearch's diagnose step), `skills/save/SKILL.md` (captures insights from autoresearch runs) ## Example output (autoresearch run on a content Critic prompt) ```markdown # autoresearch run. content Critic prompt, round 3 date: 2026-05-11 baseline_score: 7.2/10 (avg over 5 historical drafts) candidate_count: 3 winner: Critic-B (score 8.4/10, +1.2pp lift) ## diagnose (regression on round 2): - root cause: Critic prompt over-indexed on "sourcing" dimension, lost "precision" calibration - trace: 3 of 5 drafts cited correct source...