Agentic eval vs Writing lean proofs
Agentic eval is an open-source agent skill for ai models and llm tooling. Writing lean proofs is an open-source agent skill for developer tools.
- Agentic eval has more GitHub stars (41,254 vs 8,243).
- Agentic eval is best for add a self-review loop to a report generator before it returns results.
- Writing lean proofs is best for prove a theorem in lean 4 from a stated goal.
- What it is
- Agentic eval is an open-source agent skill for ai models and llm tooling.Writing lean proofs is an open-source agent skill for developer tools.
- Best for
- Add a self-review loop to a report generator before it returns resultsProve a theorem in Lean 4 from a stated goal
- Kind
- SkillSkill
- Pricing
- Open sourceOpen source
- Platforms
- Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, GooseClaude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, Goose
- GitHub stars
- 41.3K8.2K
- Last commit
- Nov 2026Nov 2026
- License
- MITCC-BY-SA-4.0
- Open source
- Good for
- Add a self-review loop to a report generator before it returns results; Build an LLM-as-judge step that ranks two candidate answers; Set up a rubric that scores agent output on accuracy and clarityProve a theorem in Lean 4 from a stated goal; Refactor a long tactic proof into smaller named lemmas; Fill in sorry placeholders in an existing Lean development
What Agentic eval does
Basic reflection loop
Self-critique against criteria with PASS/FAIL feedback and iterative refinement until all checks pass.
Evaluator-optimizer pipeline
Separates generation, scoring and optimization with a score threshold to decide when to stop.
Test-driven code refinement
Generates tests for code, runs them and feeds failures back into repeated fix attempts.
Rubric and judge evaluation
Covers outcome checks, pairwise LLM-as-judge comparison and weighted rubric scoring.
Loop safety practices
Recommends iteration caps, convergence detection, trajectory logging and structured JSON output.
What Writing lean proofs does
Top-down design workflow
Guides defining types and API lemmas before proving theorems about them.
Sorry skeleton scaffolding
Lays out statements and calc or have structures with sorry placeholders that compile first.
Focused goal handling
Enforces one goal at a time with focusing dots, show and change statements.
Lemma extraction guidance
Explains when repeated or unwieldy sub-arguments should become local haves or standalone lemmas.
Axiom-based proof gating
Uses kernel axiom collection to detect remaining sorry or extra trust assumptions.
Linter selection and custom linters
Chooses Mathlib linters by project role and writes env_linter checks for project conventions.
Agentic eval vs Writing lean proofs: questions
Is Agentic eval or Writing lean proofs cheaper?
Both are open source. Compare current plans on each official site.
What is the difference between Agentic eval and Writing lean proofs?
Agentic eval is an open-source agent skill for ai models and llm tooling. Writing lean proofs is an open-source agent skill for developer tools. Agentic eval has more GitHub stars (41,254 vs 8,243).

