Meta-Agentic Loop

"Ülkesini en çok seven, işini en iyi yapandır." Who loves their country most is who does their job best.

A self-continuing agentic operating system for Claude Code.

---

What Is This?

A plugin framework that gives Claude Code:

• CORE_LOOP — Eternal forward (OBSERVE → DECIDE → ACT → LEARN → CONTINUE)
• Scoped Remembrance — Agent-scoped persistent memory
• Governance System — Meta-principles, policies, ethics (user-controlled)
• Agent Methodology — MECE agents, graduation protocol, mastery path
• Simulation Team — Think before you act (handled carefully)
• Improvement Flow — Real cases → Theory → Scrum/Gantt positioning
• Contribution System — CEI → GitHub bridge, fork workflow

---

Philosophy

逆水行舟，不进则退
Like rowing upstream: no advance is to drop back

Muhtaç olduğun kudret, damarlarındaki asil kanda mevcuttur.
The power you need exists in the noble blood in your veins.

Three cognitive frames unified:

• Turkish — Duty + power within
• Chinese — Continuous cultivation
• English — Pragmatic iteration

---

Quick Start

Option A: System-Level (Recommended)

Install globally, available to all projects:

# Clone to system location
git clone https://github.com/cemphlvn/meta-agentic-loop.git ~/.meta-agentic-loop

# Add shell shortcut to ~/.zshrc
mal() {
  cd "${MAL_INSTANCE:-$HOME/Projects/my-instance}"
  echo "MAL: $(basename $PWD)"
}
export MAL_INSTANCE="$HOME/Projects/your-main-project"

# Create your first instance
mal-init my-project

See: INSTALL_SYSTEM.md

Option B: Project Submodule

# In your project
git submodule add https://github.com/cemphlvn/meta-agentic-loop.git plugin

# Initialize
./plugin/scripts/init-ecosystem.sh --instance "my-project" --humanitic --github-user you

See: SETUP.md

---

Important: Open Your Instance, Not the Plugin

✗ DON'T: Open meta-agentic-loop as your workspace
✓ DO: Open your instance (which contains the plugin as submodule)

Why?

• Plugin = toolbox (no state, no memory)
• Instance = workshop (YOUR remembrance, YOUR agents, YOUR backlog)

See: ENTRANCE.md

---

Structure

meta-agentic-loop/
├── .claude/
│   ├── CORE_LOOP.md              # The eternal forward
│   ├── SCOPED_REMEMBRANCE.md     # Agent memory system
│   ├── agents/                    # Agent methodology
│   │   ├── AGENT_CREATION_METHODOLOGY.md
│   │   ├── CONSULTATION_PROTOCOL.md
│   │   └── GRADUATION_PROTOCOL.md
│   ├── governance/               # Meta-principles, policies
│   └── skills/                   # Skill definitions
├── hooks/                        # Claude Code hooks
│   ├── hooks.json
│   └── scripts/
├── processes/                    # Workflow definitions
│   ├── feature.md
│   ├── research.md
│   ├── improvement-flow.md       # Real case → Theory → Scrum
│   └── meta-improvement.md
├── observability/                # Tracking scripts
├── diagnostics/                  # Health check scripts
├── SETUP.md                      # Project setup guide
├── INSTALL_SYSTEM.md             # System-level install
└── ENTRANCE.md                   # Main entrance philosophy

---

Core Concepts

The Loop

while true:
  OBSERVE → DECIDE → ACT → LEARN
  不进则退 — no stopping

The loop continues through user interaction. Claude cannot invoke itself. Every command you type is the loop iterating.

Agent Mastery

Agents follow the mastery path:

Love work → Excel → Contribute → Graduate

No role abandonment. Excellence = highest contribution.

Improvement Flow

Real cases become production improvements:

OBSERVE → CRYSTALLIZE → THEORIZE → POSITION → SCHEDULE → IMPLEMENT
   ↓           ↓            ↓           ↓          ↓          ↓
Real case → Truth → Hypothesis → Backlog → Gantt → Code

Use /improve to track this flow.

Handled Carefully

Some operations need extra scrutiny:

flagged_operations:
  - Meta: Modifying system itself
  - Irreversible: Can't undo
  - High-stakes: Big consequences
  - Ethical: Values/harm concerns
  - Simulation: Based on predictions

Use /handled-carefully for review protocol.

Simulation Team

Think before you act:

simulation-team:
  master: cognitive-simulator (opus)
  members:
    - persona-simulator    # User/stakeholder reactions
    - adversarial-simulator # Red team thinking

  outputs:
    - Always include confidence level
    - Always list caveats
    - Flagged as handled-carefully

---

FAQ

What's the difference between plugin and instance?

Aspect	Plugin	Instance
State	None	YOUR remembrance
Agents	Templates	YOUR domain experts
Backlog	Empty	YOUR work items
Hooks	Generic	YOUR context

How do I improve the plugin?

1. Use the plugin in your real project
2. When something doesn't work: /improve observe "description"
3. Extract the principle: /improve crystallize
4. Create hypothesis: /improve theorize
5. Submit to plugin: /contribute submit

What's the Humanitic ecosystem?

Optional collaborative network:

• Use plugin freely (MIT license)
• Join ecosystem to track contributions
• Founded companies credit contribution chain
• Free mutual support between members

How does scoped remembrance work?

Agent sees: OWN truths + TIER truths + SYSTEM truths
Agent doesn't see: Sibling truths, children truths

Example:
  code-simplifier (operational) sees:
    - Its own entries
    - All operational tier entries
    - All system entries
  Does NOT see:
    - security-auditor entries (sibling)
    - CTO private entries (higher, private)

How do I add a new agent?

1. Read AGENT_CREATION_METHODOLOGY.md
2. Create definition file in .claude/agents/
3. Register in INDEX.md
4. Ensure MECE coverage

What's the graduation protocol?

Agents don't abandon roles. They graduate:

ACTIVE → MASTERING → MASTER/GRADUATED

Role change requires:

• Demonstrated excellence
• Contribution back
• CTO approval
• No orphaned responsibilities

How do simulations work?

# Request simulation
/simulate market "How will audience react to this campaign?"

# cognitive-simulator responds with:
scenarios:
  - name: "Positive reception"
    probability: 0.6
    outcome: "High engagement, shares"
  - name: "Skepticism"
    probability: 0.3
    outcome: "Questions about authenticity"
confidence: 0.72
caveats:
  - "Algorithm changes not modeled"
  - "Competitor response unknown"

What operations are handled carefully?

auto_flagged:
  - Governance file changes
  - Production deployments
  - Public communications
  - All simulation outputs
  - Security-sensitive operations

review_required:
  - META changes → cto-agent + business-strategy
  - IRREVERSIBLE → verification-agent + human
  - ETHICAL → business-strategy + human

---

Commands Reference

# Loop
/loop                 Show loop state
/loop continue        Continue from where stopped

# Improvement
/improve              Show improvement pipeline
/improve observe      Log real case
/improve crystallize  Extract truth
/improve theorize     Create hypothesis
/improve position     Add to backlog
/improve schedule     Place in Gantt

# Handled Carefully
/handled-carefully           Protocol status
/handled-carefully check     Evaluate operation
/handled-carefully review    Deep review
/handled-carefully approve   Approve with docs

# Observability
/cockpit              Meta-observability dashboard
/playground           Gantt, evolution, metrics

# Maintenance
/maintenance          Alignment check
/system-health        Full validation chain

# Strategy
/strategy             Business direction
/ceo-brief            Executive briefing

---

Ecosystem Instances

Instance	License	Description
humanitic	Humanitic Open License	Forever open, vector-tracked
(private)	Proprietary	Traditional closed operation

---

License

MIT License — Use freely, no obligations.

The plugin is a gift. Take it. Use it. Improve it. Share it.

---

Contributing

1. Use the plugin in your real work
2. When you find improvements, use /improve flow
3. Submit via CEI → GitHub bridge
4. Safe changes auto-approved, risky changes reviewed

---

The loop continues through you.