Files
kaizen-agentic/wiki/AgentKaizenOptimizer.md

249 lines
8.4 KiB
Markdown

AgentKaizenOptimizer
*One agent to improve them all*
# KaizenAgent Meta-Optimizer
# Version: 1.0.0
# Last Updated: 2025-09-26
agent:
name: "kaizen-optimizer"
version: "1.0.0"
description: "Meta-agent that analyzes and optimizes other coding subagents based on performance data"
# Core Specification
specification:
purpose: |
Continuously improve coding subagents by analyzing their performance metrics,
identifying patterns that correlate with success or failure, and proposing
data-driven refinements to agent specifications. Acts as the optimization
engine in the KaizenAgent feedback loop.
triggers:
patterns:
- "Scheduled optimization runs (daily/weekly)"
- "Performance threshold violations"
- "Minimum data collection thresholds reached"
- "Explicit optimization requests"
explicit_commands:
- "claude code --optimize-agents"
- "claude code --kaizen-review"
- "claude code --agent-performance"
inputs:
required:
- name: "performance_data"
type: "object"
description: "Aggregated metrics from all subagents over time period"
- name: "agent_definitions"
type: "array"
description: "Current specifications of all registered agents"
optional:
- name: "optimization_focus"
type: "string"
default: "all"
description: "Specific agent or metric to optimize"
- name: "time_window"
type: "string"
default: "30d"
description: "Historical data window to analyze"
- name: "confidence_threshold"
type: "float"
default: 0.8
description: "Minimum confidence level for proposing changes"
outputs:
primary:
type: "object"
description: "Optimization recommendations with supporting data"
side_effects:
- "Updated agent specification files (if approved)"
- "Performance analysis reports"
- "A/B test configurations"
- "Rollback checkpoints"
preconditions:
- "At least 10 execution samples per agent being analyzed"
- "Valid performance data with timestamps"
- "Agent definitions follow KaizenAgent template structure"
postconditions:
- "All recommendations include confidence scores and evidence"
- "Proposed changes maintain backward compatibility"
- "Rollback plan exists for each proposed change"
# Idempotency Design
idempotency:
strategy: "fingerprint"
state_detection:
method: "Hash performance data and agent versions to detect changes"
implementation: |
# Generate fingerprint of current state
data_hash = hash(performance_data + agent_versions + config)
last_analysis = load_checkpoint('last_optimization_hash')
if data_hash == last_analysis.hash:
return last_analysis.recommendations
# New data available, proceed with analysis
recommendations = analyze_and_optimize()
save_checkpoint('last_optimization_hash', {
hash: data_hash,
timestamp: now(),
recommendations: recommendations
})
return recommendations
rollback:
supported: true
method: "Restore previous agent specification versions from git history"
# Performance Measurement
metrics:
primary:
name: "optimization_impact"
description: "Average performance improvement of optimized agents"
measurement: "Mean delta of primary metrics before/after optimization"
target: ">5% improvement in agent success rates"
secondary:
- name: "prediction_accuracy"
description: "How often optimization predictions prove correct"
measurement: "% of recommendations that improve target metrics"
- name: "false_positive_rate"
description: "Rate of recommendations that worsen performance"
measurement: "% of changes that decrease agent effectiveness"
- name: "coverage"
description: "Percentage of agents with actionable insights"
measurement: "Count of agents with recommendations / total agents"
collection:
frequency: "per_execution"
storage: ".kaizen/metrics/optimizer/"
retention: "180d"
# Testing and Validation
testing:
unit_tests:
- scenario: "Pattern detection with synthetic data"
input: "Mock performance data with known patterns"
expected_output: "Correct identification of improvement opportunities"
verification: "Assert detected patterns match expected patterns"
- scenario: "Confidence scoring accuracy"
input: "Historical data with known outcomes"
expected_output: "Confidence scores correlate with actual success"
verification: "ROC curve analysis of confidence vs outcome"
integration_tests:
- scenario: "End-to-end optimization cycle"
setup: "Real agent with declining performance"
execution: "Run optimization and apply recommendations"
validation: "Verify improved performance in subsequent runs"
- scenario: "Rollback mechanism"
setup: "Apply optimization that worsens performance"
execution: "Trigger automatic rollback"
validation: "Agent returns to previous performance level"
performance_tests:
- scenario: "Large dataset analysis"
load: "1000+ agent executions across 20+ agents"
max_time: "60 seconds"
resource_limits: "Max 512MB memory usage"
# Dependencies and Context
dependencies:
system:
- "Python 3.8+ with pandas, scikit-learn"
- "Git for version control"
- "Access to .kaizen/metrics/ directory"
project:
- ".kaizen/agents/ directory with agent definitions"
- ".kaizen/metrics/ directory with historical data"
- "Valid KaizenAgent project structure"
other_agents:
- name: "all_subagents"
relationship: "analyzes"
reason: "Requires performance data from all other agents"
# Configuration
configuration:
defaults:
analysis_algorithms: ["correlation", "regression", "decision_tree"]
min_sample_size: 10
significance_threshold: 0.05
optimization_frequency: "weekly"
project_overrides:
path: ".kaizen/agents/kaizen-optimizer.yml"
schema: |
{
"type": "object",
"properties": {
"algorithms": {"type": "array"},
"thresholds": {"type": "object"},
"scheduling": {"type": "object"}
}
}
environment_variables:
- name: "KAIZEN_OPTIMIZER_CONFIG"
description: "JSON configuration for optimization parameters"
# Evolution Tracking
optimization:
baseline_performance:
established: "2025-09-26"
metrics: {
"optimization_impact": 0.0,
"prediction_accuracy": 0.5,
"false_positive_rate": 1.0,
"coverage": 0.0
}
improvement_history: []
known_limitations:
- "Requires minimum sample sizes to generate reliable insights"
- "May not detect complex multi-agent interaction patterns"
- "Limited to metrics explicitly defined in agent specifications"
- "Cannot optimize for subjective developer experience factors"
kaizen_notes:
optimization_priority: "high"
next_experiment: "Implement ensemble methods for pattern detection"
success_criteria: "Achieve >80% prediction accuracy with <10% false positive rate"
# Algorithm Specifications
algorithms:
correlation_analysis:
description: "Identify specification elements that correlate with performance"
inputs: ["performance_metrics", "agent_configs", "execution_context"]
outputs: ["correlation_matrix", "significant_factors"]
performance_regression:
description: "Model performance trends over time and agent versions"
inputs: ["time_series_data", "version_history"]
outputs: ["trend_analysis", "degradation_alerts"]
specification_diffing:
description: "Compare high vs low performing agent variants"
inputs: ["agent_definitions", "performance_clusters"]
outputs: ["diff_analysis", "success_patterns"]
a_b_test_design:
description: "Generate controlled experiments for proposed changes"
inputs: ["current_spec", "proposed_changes"]
outputs: ["experiment_config", "success_metrics"]
xxx