## Architecture

kaizen-agentic has two distinct layers:

### 1. Python framework (`src/kaizen_agentic/`)

- **`core.py`** — `Agent` (abstract base) + `AgentConfig` (dataclass). Tracks performance, supports config updates, implements kaizen interface.
- **`optimization.py`** — `OptimizationLoop` (runs improvement cycles, detects trends, generates recommendations) + `PerformanceMetrics` (execution time, success rate, quality scores).
- **`metrics.py`** — `MetricsStore` + `OptimizerStore` (project-scoped `.kaizen/metrics/` per ADR-004).

### 2. Agent definitions (`agents/` — 20 files)

Markdown instruction sets read and followed by Claude. Not executables. Naming convention: `agent-{name}.md`.
Packaged copies live in `src/kaizen_agentic/data/agents/` for `pip install` distribution.

| Category | Agents |
|----------|--------|
| Testing | `tdd-workflow`, `test-maintenance`, `testing-efficiency` |
| Quality | `code-refactoring`, `datamodel-optimization` |
| Process | `requirements-engineering`, `keepaTodofile`, `keepaChangelog`, `keepaContributingfile`, `project-management`, `priority-evaluation`, `scope-analyst` |
| Infrastructure | `setupRepository`, `tooling-optimization`, `sys-medic` |
| Release | `releaseManager` |
| Docs | `claude-documentation` |
| Support | `wisdom-encouragement` |
| Meta | `coach`, `optimization` |

### Custodian integration

The state-hub MCP resolves the agents directory via `host_paths[hostname]` → `local_path`. Tools: `list_kaizen_agents(category?)`, `get_kaizen_agent(name)`.