harness-optimizer

⚡ Harness Optimizer — Tune Your Agent for Peak Performance

Philosophy: The best agent configuration is the one that ships the most correct code per dollar spent. Optimize for outcome-per-token, not raw capability.

1. When to Use This Skill

• Agent token costs are too high
• Agent is hitting context window limits
• Agent performance is inconsistent or degrading
• Switching between models for different tasks
• Setting up a new project's agent configuration
• Debugging why agent quality has dropped

2. Token Cost Optimization

The Token Budget Framework

Task Type	Recommended Model	Context Budget	Target Cost
Simple code edits	Fast/small model	8K tokens	$0.01-0.05
Code review	Medium model	16K tokens	$0.05-0.20
Architecture design	Best available model	32K tokens	$0.20-1.00
Complex debugging	Best available model	64K tokens	$0.50-2.00
Documentation	Medium model	16K tokens	$0.05-0.15
Test generation	Medium model	32K tokens	$0.10-0.30

Cost Reduction Strategies

Strategy 1: Tiered Model Routing

# Route tasks to the cheapest model that can handle them
routing:
  simple_edits:
    model: claude-3-haiku     # Cheapest
    max_tokens: 4096
  code_review:
    model: claude-3.5-sonnet  # Mid-tier
    max_tokens: 8192
  architecture:
    model: claude-3.5-opus    # Premium — only when needed
    max_tokens: 16384

Strategy 2: Context Compression

## Before (wasteful — 2000 tokens)
Here is the entire file contents of user_service.py:
[... 200 lines of code ...]
Please fix the bug on line 45.

## After (efficient — 400 tokens)
In user_service.py, the `create_user` method (lines 40-55):
```python
def create_user(self, data):
    # BUG: missing email validation
    user = User(**data)
    self.db.save(user)

Fix: add email validation before line 43.

#### Strategy 3: Result Caching
```python
# Cache expensive operations
# Don't re-analyze unchanged files
cache = load_cache(".agent/context/recall_cache.json")

if file_hash == cache.get(file_path, {}).get("hash"):
    # File hasn't changed — reuse previous analysis
    return cache[file_path]["result"]
else:
    # File changed — re-analyze
    result = analyze(file_path)
    cache[file_path] = {"hash": file_hash, "result": result}
    save_cache(cache)