kaizen-agentic/docs/agency-framework.md

# Agency Framework

kaizen-agentic is not just a library of agent instruction sets — it is an **agency**: a system where agents are deployed into projects with their own persistent memory, learn from experience, and are guided by a coaching meta-agent that distils patterns across the entire fleet.

## Overview

When you deploy a kaizen-agentic agent into a project, it can accumulate **project-scoped memory** — a structured file written at session close and read at session start. A **Coach** meta-agent reads across all agent memories and produces orientation briefs for newly deployed agents: what has been tried, what worked, what to watch out for.

Agents arrive in a project informed, not blank.

---

## Project Memory

### Location Convention

```
<project-root>/.kaizen/agents/<agent-name>/memory.md
```

The `.kaizen/` directory is analogous to `.claude/` — a project-level configuration and state directory owned by the kaizen-agentic ecosystem.

### Memory File Structure

```markdown
---
agent: <name>
project: <project-root or slug>
last_updated: <ISO date>
session_count: <n>
---

## Project Context
<!-- What this agent knows about the project it is working in -->

## Accumulated Findings
<!-- Patterns, recurring issues, key decisions the agent has encountered -->

## What Worked
<!-- Approaches that produced good results in this project -->

## Watch Points
<!-- Recurring risks, traps, or areas requiring extra care -->

## Open Threads
<!-- Things noticed but not yet acted on -->

## Session Log
<!-- One-line entry per session: date · summary · outcome -->
```

### Session Protocols

**Session-start (all agents with `memory: enabled`):**

1. Check for `.kaizen/agents/<name>/memory.md` in the project root.
2. If present, read it before beginning work.
3. Acknowledge the memory in the opening brief.

**Session-close (all agents with `memory: enabled`):**

1. Update `## Accumulated Findings`, `## What Worked`, `## Watch Points` as appropriate.
2. Append one line to `## Session Log`: `YYYY-MM-DD · <summary> · <outcome>`.
3. Bump `last_updated` and `session_count`.

---

## Agent YAML Frontmatter

Each agent definition (`agents/agent-<name>.md`) includes a YAML frontmatter block:

```yaml
---
name: <name>
description: <one-line description>
category: <category>
memory: enabled   # or: disabled
---
```

The `memory` field defaults to `enabled`. Set `memory: disabled` for agents that are stateless by design (e.g. `wisdom-encouragement`).

---

## The Coach Meta-Agent

`agents/agent-coach.md` is a **meta-agent** — it performs no domain work (coding, testing, infrastructure). Its sole purpose is synthesis and advice.

### What the Coach Does

- **Cross-agent synthesis**: reads all `.kaizen/agents/*/memory.md` files, identifies shared patterns, cross-domain risks, and contradictions
- **New-agent orientation**: when briefing a specific agent, filters all existing memories for what is relevant and produces a targeted brief
- **Fleet health overview**: summarises which agents are active, stale, or missing; flags high-session-count agents with open threads

### Invoking the Coach

**Via CLI (assembles raw memory context):**

```bash
kaizen-agentic memory brief <agent-name>
```

This prints a structured orientation brief. Pass the output to a Claude session with `agents/agent-coach.md` loaded for full LLM synthesis.

**Directly in a Claude session:**

```
Coach, brief the sys-medic agent on this project.
Coach, what patterns have you observed across all agents?
```

The Coach maintains its own memory at `.kaizen/agents/coach/memory.md` covering fleet-level observations over time.

---

## CLI Reference

The `memory` command group manages project-scoped agent memory:

```
kaizen-agentic memory show <agent>      # Print agent memory for the current project
kaizen-agentic memory init <agent>      # Scaffold an empty memory file
kaizen-agentic memory brief <agent>     # Assemble orientation context for the coach
kaizen-agentic memory clear <agent>     # Wipe memory (with confirmation prompt)
```

### Options

`memory brief` accepts:
- `--target / -t` — project root (default: current directory)
- `--raw` — dump raw memory file contents without the structured header

### Example Workflow

```bash
# First deployment of sys-medic into a project
kaizen-agentic memory init sys-medic

# After a few sessions, brief an incoming tdd-workflow agent
kaizen-agentic memory brief tdd-workflow
# → paste output into Claude with agent-coach.md loaded

# Review accumulated memory for a specific agent
kaizen-agentic memory show project-management
```

---

## Protocols (Part 3 — coming in WP-0002 Part 3)

A future extension adds **protocol runbooks** — structured, human-readable procedural checklists that agents can reference during structured assessments:

```
agents/protocols/<agent-name>/<slug>.md
```

The sys-medic k3s health assessment protocol is the first planned example. CLI commands `kaizen-agentic protocols list` and `protocols show` will expose them.

See [WP-0002](../workplans/kaizen-agentic-WP-0002-agency-framework.md) Part 3 for the full specification.

---

## Agents with Memory Enabled

All agents that do session-bound project work have `memory: enabled` in their frontmatter and include session-start/session-close protocol blocks:

| Agent | Category | Notes |
|-------|----------|-------|
| project-management | process | Reference implementation of the session protocol pattern |
| tdd-workflow | testing | |
| requirements-engineering | process | |
| scope-analyst | process | |
| sys-medic | infrastructure | Extended memory template (node profiles, recurring findings) |
| coach | meta | Fleet-level memory |

---

## Related Documents

- [ADR-001: Workplan Convention](../workplans/kaizen-agentic-WP-0001-community-engagement.md) — how work items are structured
- [ADR-002: Project Memory Convention](../workplans/kaizen-agentic-WP-0002-agency-framework.md) — memory file location, structure, and lifecycle
- [WP-0002: Agency Framework](../workplans/kaizen-agentic-WP-0002-agency-framework.md) — full implementation workplan