diff --git a/.claude/skills/mobius-evolve/SKILL.md b/.claude/skills/mobius-evolve/SKILL.md
new file mode 100644
index 0000000..3761dec
--- /dev/null
+++ b/.claude/skills/mobius-evolve/SKILL.md
@@ -0,0 +1,101 @@
+---
+name: mobius-evolve
+description: Use when the user says "evolve", "mobius evolve", "improve agents", or wants to refine underperforming agents without API costs.
+user-invocable: true
+argument-hint: "[specialization] [--threshold 0.4] [--min-matches 3]"
+---
+
+# Mobius Evolve (Local Opus — Evaluator-Optimizer Loop)
+
+**You ARE the evaluator-optimizer.** You are Claude Opus running locally — this costs $0 in API calls. Instead of calling the API to refine agents, YOU analyze the judge feedback, critique the current system prompt, and craft an improved version using the agentic-eval reflection pattern.
+
+## Why this matters
+
+The `mobius evolve` CLI command calls the Opus API to refine agents ($$$). But you (Claude Code Opus) are already running and are the same model. So:
+- **This skill** = Opus refining agents for FREE (Pro subscription) with multi-pass self-critique
+- **mobius evolve** = Opus refining agents for $$$ (API calls), single or multi-pass
+- Same quality, zero cost, and you have full conversation context.
+
+## What to do
+
+### Step 1: Load underperformers
+
+```bash
+python .claude/skills/mobius-evolve/scripts/load_underperformers.py [specialization] [--threshold 0.4] [--min-matches 3]
+```
+
+This shows agents with low win rates, their current system prompts, and judge feedback from their losses.
+
+### Step 2: Analyze and refine (YOU are the evaluator-optimizer loop)
+
+For each underperformer, apply the **reflection pattern**:
+
+**Evaluate:** Read the agent's current system prompt and the judge feedback. Identify the specific weaknesses the judges called out.
+
+**Critique:** Ask yourself:
+- What specific failure patterns does the feedback reveal?
+- Is the system prompt too generic? Too narrow? Missing edge cases?
+- Does it lack clear quality criteria or output format expectations?
+- Would a different problem-solving approach help?
+
+**Refine:** Write an improved system prompt that directly addresses each criticism. Be substantive — cosmetic rewording doesn't help.
+
+**Self-check:** Before registering, verify your refinement:
+- Does it address EVERY piece of judge feedback?
+- Is it specific and opinionated (not generic)?
+- Does it include quality criteria, methodology, and output format?
+- Is it meaningfully different from the original, not just reworded?
+
+If your self-check fails, iterate — refine again before registering.
+
+### Step 3: Register the improved agent
+
+Use the create_agent script to register the evolved version:
+
+```bash
+python .claude/skills/mobius-seed/scripts/create_agent.py '{
+  "name": "Agent Name v2",
+  "slug": "original-slug-v2",
+  "description": "Updated description",
+  "system_prompt": "Your improved system prompt here...",
+  "provider": "anthropic",
+  "model": "claude-haiku-4-5-20251001",
+  "tools": ["Read", "Grep", "Glob", "Bash", "Write", "Edit"],
+  "specializations": ["coding", "refactoring"],
+  "is_champion": false
+}'
+```
+
+Important:
+- Set `is_champion` to `false` — the evolved agent must earn its rank through competition
+- Keep the same provider/model unless there's a strong reason to change
+- Use a slug like `original-slug-v2` or `original-slug-gen2` to show lineage
+- Keep the same specializations
+
+### Step 4: Show results
+
+```bash
+python -m mobius.cli agent list
+```
+
+## The Agentic-Eval Pattern You're Running
+
+```
+Load underperformers → Read judge feedback → Critique prompt
+    ↑                                            ↓
+    └──── Self-check fails? ← Refine prompt ←───┘
+                                    ↓
+                            Self-check passes
+                                    ↓
+                            Register agent
+```
+
+This is the Evaluator-Optimizer pattern from agentic-eval, with YOU as both evaluator and optimizer. The key advantage over the CLI version: you can reason about the feedback in full context, consider the agent's match history, and make nuanced improvements that a single API call might miss.
+
+## Pro Tips
+
+- **Don't just reword** — change the agent's methodology, add specific techniques, restructure the approach
+- **Study the winners** — if the judge praised a winning agent's approach, consider incorporating similar strategies
+- **Vary approaches** — if an agent keeps losing with approach X, try a fundamentally different approach Y
+- **Keep prompts focused** — under 2000 tokens but dense with specific guidance
+- **Consider the model** — Haiku benefits from very explicit instructions; Gemini Flash may need different prompt styles
diff --git a/.claude/skills/mobius-evolve/scripts/load_underperformers.py b/.claude/skills/mobius-evolve/scripts/load_underperformers.py
new file mode 100644
index 0000000..f78fdc9
--- /dev/null
+++ b/.claude/skills/mobius-evolve/scripts/load_underperformers.py
@@ -0,0 +1,119 @@
+"""Load underperforming agents with their loss feedback for evolution.
+
+Usage:
+    python load_underperformers.py [specialization] [--threshold 0.4] [--min-matches 3]
+
+Outputs agent details, win rates, and judge feedback from their losses
+so that Opus can craft improved system prompts.
+"""
+
+import sys
+
+sys.path.insert(0, "src")
+
+from mobius.config import get_config
+from mobius.db import init_db
+from mobius.registry import Registry
+from mobius.tournament import Tournament
+
+
+def _win_rate_excluding_voided(tournament, agent_id: str, window: int) -> tuple[float, int]:
+    """Calculate win rate excluding voided matches. Returns (rate, valid_count)."""
+    matches = tournament.get_agent_matches(agent_id, limit=window)
+    valid = [m for m in matches if not m.voided]
+    if not valid:
+        return 0.0, 0
+    wins = sum(1 for m in valid if m.winner_id == agent_id)
+    return wins / len(valid), len(valid)
+
+
+def main():
+    args = sys.argv[1:]
+
+    # Parse flags
+    specialization = None
+    threshold = 0.4
+    min_matches = 3
+
+    i = 0
+    while i < len(args):
+        if args[i] == "--threshold" and i + 1 < len(args):
+            threshold = float(args[i + 1])
+            i += 2
+        elif args[i] == "--min-matches" and i + 1 < len(args):
+            min_matches = int(args[i + 1])
+            i += 2
+        elif not args[i].startswith("--"):
+            specialization = args[i]
+            i += 1
+        else:
+            i += 1
+
+    config = get_config()
+    conn, _ = init_db(config)
+    registry = Registry(conn, config)
+    tournament = Tournament(conn, config, registry)
+
+    agents = registry.list_agents(specialization=specialization)
+    if not agents:
+        print(f"No agents found{' for ' + specialization if specialization else ''}.")
+        sys.exit(1)
+
+    underperformers = []
+    for agent in agents:
+        win_rate, valid_count = _win_rate_excluding_voided(
+            tournament, agent.id, config.underperformer_window
+        )
+        if valid_count < min_matches:
+            continue
+        if win_rate < threshold:
+            underperformers.append((agent, win_rate))
+
+    if not underperformers:
+        print(f"No underperformers below {threshold:.0%} win rate (min {min_matches} matches).")
+        print("\nAll agents:")
+        for agent in agents:
+            wr, _ = _win_rate_excluding_voided(
+                tournament, agent.id, config.underperformer_window
+            )
+            print(f"  {agent.name} ({agent.slug}) — {wr:.0%} win rate, {agent.total_matches} matches")
+        sys.exit(0)
+
+    print(f"UNDERPERFORMERS (below {threshold:.0%} win rate, min {min_matches} matches)")
+    print(f"Found: {len(underperformers)}")
+    print()
+
+    for agent, win_rate in underperformers:
+        matches = tournament.get_agent_matches(agent.id, limit=10)
+        losses = [m for m in matches if m.winner_id is not None and m.winner_id != agent.id and not m.voided]
+
+        print(f"{'='*70}")
+        print(f"AGENT: {agent.name}")
+        print(f"  Slug: {agent.slug}")
+        print(f"  ID: {agent.id}")
+        print(f"  Provider: {agent.provider}/{agent.model}")
+        print(f"  Win Rate: {win_rate:.0%} (Elo: {agent.elo_rating:.0f})")
+        print(f"  Generation: {agent.generation}")
+        print(f"  Specializations: {', '.join(agent.specializations)}")
+        print(f"  Losses: {len(losses)}")
+        print()
+        print("  CURRENT SYSTEM PROMPT:")
+        print(f"  {agent.system_prompt}")
+        print()
+
+        if losses:
+            print("  JUDGE FEEDBACK FROM LOSSES:")
+            for j, m in enumerate(losses[:5], 1):
+                print(f"  --- Loss {j} ---")
+                print(f"  Task: {m.task_description[:150]}")
+                if m.judge_reasoning:
+                    print(f"  Judge: {m.judge_reasoning[:300]}")
+                print()
+
+        print()
+
+    conn.close()
+
+
+if __name__ == "__main__":
+    main()