Add raphaelchristi/harness-evolver

agents/raphaelchristi__harness-evolver/README.md

@@ -0,0 +1,68 @@
+# Harness Evolver
+
+**Harness Evolver** is a LangSmith-native autonomous agent optimizer for Claude Code. Point it at any LLM-based agent codebase and it will iteratively improve prompts, routing, tools, and architecture — automatically, with full evaluation-backed evidence for every change.
+
+## What It Does
+
+Harness Evolver runs an automated evolution loop on your AI agent:
+
+1. **Sets up a LangSmith evaluation dataset** from your existing traces or test cases
+2. **Spawns self-organizing proposer agents**, each investigating a data-driven lens (failing example patterns, routing issues, tool gaps)
+3. **Tests each proposal** in an isolated git worktree — no risk to your main branch
+4. **Evaluates with LangSmith** using LLM-as-judge with rubrics, pairwise comparison, and justification-before-score
+5. **Gates regressions** — constraint checks, efficiency gates, Pareto selection, holdout enforcement
+6. **Merges only winners** and builds cross-iteration memory to guide future proposals
+
+## Key Capabilities
+
+- **Multi-agent orchestration**: 6 specialized sub-agents (proposer, evaluator, critic, architect, consolidator, testgen)
+- **30+ Python tools** for LangSmith integration, trace analysis, architecture analysis, regression tracking, and more
+- **Evolution archive**: TF-IDF search over all past candidates (winners and losers) to avoid repeating failures
+- **Self-abstention**: proposers honestly abstain when they can't add value
+- **Secret detection**: all outputs filtered for API keys before logging
+- **Compound learning**: proven evolution learnings are promoted back into CLAUDE.md
+
+## Install
+
+```bash
+# Via Claude Code plugin marketplace (recommended)
+/plugin marketplace add raphaelchristi/harness-evolver-marketplace
+/plugin install harness-evolver
+
+# Or via npx
+npx harness-evolver@latest
+```
+
+## Quick Start
+
+```bash
+cd my-llm-project
+export LANGSMITH_API_KEY="lsv2_pt_..."
+claude
+
+/harness:setup      # explores project, configures LangSmith
+/harness:health     # check dataset quality
+/harness:evolve     # runs the optimization loop
+/harness:status     # check progress (rich ASCII chart)
+/harness:deploy     # tag, push, finalize
+```
+
+## Real Results
+
+Tested on a RAG agent (Agno framework, Gemini Flash Lite):
+
+| Iteration | Score | Action |
+|---|---|---|
+| baseline | 0.575 | Original agent with hallucinations and broken tool calls |
+| v002 | 0.950 | Breakthrough: inlined KB into prompt, eliminated vector search (5.7x faster) |
+| v007 | 1.000 | One-shot example injection + rubric-aligned responses — perfect on held-out |
+
+## Works With
+
+Claude Code, Cursor, Codex, Windsurf
+
+## Links
+
+- [GitHub](https://github.com/raphaelchristi/harness-evolver)
+- [npm](https://www.npmjs.com/package/harness-evolver)
+- [Meta-Harness paper](https://arxiv.org/abs/2603.28052)

agents/raphaelchristi__harness-evolver/metadata.json

@@ -0,0 +1,15 @@
+{
+  "name": "harness-evolver",
+  "author": "raphaelchristi",
+  "description": "Autonomous LLM agent optimizer for Claude Code. Evolves agent prompts, tools & architecture using multi-agent proposers, LangSmith evaluation, and git worktrees.",
+  "repository": "https://github.com/raphaelchristi/harness-evolver",
+  "path": "",
+  "version": "6.4.2",
+  "category": "developer-tools",
+  "tags": ["agent-optimization", "langsmith", "claude-code", "evolution", "meta-harness", "llm", "evaluation"],
+  "license": "MIT",
+  "model": "claude-sonnet-4-5-20250929",
+  "adapters": ["claude-code", "system-prompt"],
+  "icon": false,
+  "banner": false
+}