Existence file

skills-autoresearch-flue

A Flue-based research harness for deciding whether agent skills help, how much context they should contain, and when a model no longer needs them. It exists to make skill changes evidence-driven instead of vibes-driven.

Goal

Run auditable eval loops across producer models, compare no-skill and skill-guided output, inspect scores and costs, and let a researcher agent propose the smallest useful skill change when guidance is justified.

  • TypeScript
  • Agents
  • Evaluation

Now

Where it is

The alpha centers on a release-notes fixture with baseline import, optional research, producer evals, independent judging, score aggregation, cost summaries, and persisted artifacts.

Next

What comes next

  • Tighten resume and retry behavior without overwriting audit evidence.
  • Keep documentation, examples, schemas, and fixture behavior aligned as the alpha loop changes.
  • Explore cross-provider judging after the single-provider flow is stable enough for meaningful comparisons.