Meta‑Agentic α‑AGI 👁️✨ Demo – Production‑Grade v0.1.0

preview

Current runnable path — 1.14.0

Mode: Synthetic evaluation. Runs provider-driven code proposals, synthetic fitness and SQLite lineage.

Prerequisites: Python 3.11–3.13; installed project dependencies.

python -m alpha_factory_v1.demos check meta_agentic_agi
python -m alpha_factory_v1.demos run meta_agentic_agi

The catalog command uses bundled inputs and explicit offline defaults.

Expected result: Three generations saved to lineage.sqlite; repeat runs append safely.

Scope: The mock provider and pseudo-accuracy are explicitly synthetic, not measured task performance.

The catalog explains installation, stopping, backups and recovery. Browser charts for legacy demos are labeled sample replays. Original research narratives and advanced scripts below are preserved; they do not expand the tested scope stated here.

This repository is a conceptual research prototype. References to "AGI" and "superintelligence" describe aspirational goals and do not indicate the presence of a real general intelligence. Use at your own risk. Nothing herein constitutes financial advice. MontrealAI and the maintainers accept no liability for losses incurred from using this software. Each demo package exposes its own __version__ constant. The value marks the revision of that demo only and does not reflect the overall Alpha‑Factory release version.

Official definition – Meta-Agentic (adj.)
Describes an agent whose primary role is to create, select, evaluate, or re‑configure other agents and the rules governing their interactions, thereby exercising second‑order agency over a population of first‑order agents.

The term was pioneered by Vincent Boucher, President of MONTREAL.AI.

%% 𝗚𝗿𝗮𝗻𝗱 𝗦𝘆𝗻𝗮𝗽𝘀𝗲 𝗚𝗿𝗮𝗽𝗵 – Meta‑Agentic α‑AGI
graph LR
  classDef meta fill:#6425ff,stroke:#eee,color:#fff
  classDef layer fill:#1e1e2e,stroke:#ddd,color:#fff
  classDef agent fill:#0f9d58,stroke:#fff,color:#fff
  classDef tool  fill:#fbbc05,stroke:#000,color:#000
  %% Layers
  A0["🧠 Meta‑Programmer"]:::meta
  A1["📈 Evolution Archive"]:::layer
  A2["⚖️ Multi‑Objective Scorer"]:::layer
  A3["🧩 Agent Population"]:::layer
  %% Agents & Tools
  subgraph " "
    direction TB
    D1["🔍 Researcher"]:::agent
    D2["👷 Builder"]:::agent
    D3["🧪 Evaluator"]:::agent
    D4["🛠 Auto‑Tuner"]:::agent
    D5["🛡 Guardian"]:::agent
  end
  subgraph " "
    direction TB
    T1["GPT‑4o"]:::tool
    T2["Claude‑3"]:::tool
    T3["Llama‑3 ∞"]:::tool
  end
  %% Links
  A0 -->|generate| A3
  A3 -->|select| A2
  A2 -->|rank| A1
  A1 -- feedback --> A0
  %% Providers
  D1 -.uses.-> T1
  D2 -.uses.-> T3
  D3 -.uses.-> T2
  D4 -.uses.-> T3
  D5 -.uses.-> T1
  %% Value loop
  subgraph " "
    direction LR
    V1["🌐 Industry Data Streams"]
    V2["💎 Extracted Alpha"]
    V3["🚀 Deployed Solutions"]
  end
  A3 -->|iterate| V1
  V1 -->|signals| D1
  D2 --> V2
  V2 --> D3
  D4 --> V3
  D5 -.audit.-> V3

Elevating Alpha‑Factory v1 into a self‑improving, cross‑industry “Alpha Factory” that systematically
Out‑Learn · Out‑Think · Out‑Design · Out‑Strategize · Out‑Execute — without coupling to a single vendor or model.
Inspired by and extending the Meta‑Agent Search paradigm from Hu et al. (ICLR 2025).

📌 Purpose & Positioning

This demo operationalises the Automated Design of Agentic Systems (ADAS) paradigm and layers:

True multi‑objective search (accuracy, cost, latency, risk, carbon)
Open‑weights or API‑based FM back‑ends (OpenAI, Anthropic, Mistral .gguf …)
Automated provenance & lineage visualisation
Antifragile, regulator‑ready safeguards

into the existing Alpha‑Factory v1 (multi‑agent AGENTIC α‑AGI) pipeline.

1 Quick‑start 🏁

Minimal install

git clone https://github.com/MontrealAI/AGI-Alpha-Agent-v0.git
cd AGI-Alpha-Agent-v0/alpha_factory_v1/demos/meta_agentic_agi

python3 -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# Optional sanity check – verifies optional extras and offers
# an offline install flow when `WHEELHOUSE` points to a local
# directory of wheels.
python ../../check_env.py

python app.py --gens 6 --provider mistral:7b-instruct.gguf --ui
# UI → http://localhost:8501

Micromamba workflow

micromamba create -n metaagi python=3.11 -y
micromamba activate metaagi
pip install -r requirements.txt        # ≤ 40 MiB wheels

python meta_agentic_agi_demo.py --provider mistral:7b-instruct.gguf
# …or use OPENAI_API_KEY=sk‑… python meta_agentic_agi_demo.py --provider openai:gpt-4o
streamlit run ui/lineage_app.py

meta_agentic_search/sample_task.json includes a tiny ARC example used by the demo. The path resolves automatically when running the scripts.

No GPU? llama‑cpp‑python auto‑selects 4‑bit quantisation < 6 GB RAM.

🎓 Colab notebook

Spin up the demo end‑to‑end without installing anything. Works offline using open‑weights or with your API keys. The notebook now includes a quick smoke test cell to verify the demo setup.

🤖 OpenAI Agents & ADK

Use openai_agents_bridge.py to expose the search loop via the OpenAI Agents SDK:

python openai_agents_bridge.py
# → http://localhost:5001/v1/agents

For cross-process federation set ALPHA_FACTORY_ENABLE_ADK=true to auto-register agents with the bundled Google ADK gateway (alpha_factory_v1/backend/adk_bridge.py).

2 Folder Structure 📁

meta_agentic_agi/
├── core/                # provider‑agnostic primitives
│   ├── fm.py            # unified FM wrapper
│   ├── prompts.py       # reusable prompt fragments
│   └── tools.py         # exec sandbox, RAG, vector store
├── meta_agentic_search/ # ⬅ evolutionary loop
│   ├── archive.py       # stepping‑stone JSONL log
│   ├── search.py        # NSGA‑II + Reflexion
│   └── scorer.py        # multi‑objective metrics
├── agents/
│   ├── agent_base.py    # runtime interface
│   └── seeds.py         # bootstrap population
├── ui/
│   ├── lineage_app.py   # Streamlit dashboard
│   └── assets/
├── configs/
│   └── default.yml      # editable in‑UI
└── meta_agentic_agi_demo.py

3 High‑Level Architecture 🔍

graph TD
  subgraph "Meta Agent Search Loop"
    MGPT["Meta LLM Programmer"]
    Candidate["Candidate Agent<br/>Python fn"]
    Evaluator["Sandboxed Evaluator"]
    Archive["Archive<br/>(Pareto + Novelty)"]
    MGPT -->|generates| Candidate
    Candidate --> Evaluator
    Evaluator -->|scores| Archive
    Archive -->|context & feedback| MGPT
  end
  UI["Streamlit Lineage UI"] <-->|stream
lineage| Archive

flowchart LR
  AFV1["Alpha‑Factory v1 Core"]
  MAA["Meta‑Agentic Layer"]
  Providers["FM Providers<br/>(OpenAI / Anthropic / llama‑cpp)"]
  Dataset["Domain Datasets"]
  UI2["Lineage UI"]
  AFV1 --> MAA
  MAA --> Providers
  MAA --> Dataset
  Archive -.-> UI2

4 Provider Abstraction ➡️ open‑weights 🏋️‍♀️

configs/default.yml (excerpt):

provider: mistral:7b-instruct.gguf   # any ollama / llama.cpp id
context_length: 8192
rate_limit_tps: 4
retry_backoff: 2

Change provider to:

Value	Notes
`openai:gpt-4o`	needs `OPENAI_API_KEY`
`anthropic:claude-3-sonnet`	needs `ANTHROPIC_API_KEY`
`mistral:7b-instruct.gguf`	default local model

Wrapper normalises chat/completions, streams via MCP, and window‑slides tokens.

5 Multi‑Objective Search 🎯

Objective vector = [accuracy, cost, latency, hallucination‑risk, carbon]

NSGA‑II elitist selection
Behaviour descriptor = SHA‑256 of candidate AST
Optional human‑in‑the‑loop thumbs up/down (UI)

6 Security & Antifragility 🛡

Firejail --seccomp + 512 MiB mem‑cgroup sandbox
Static analysis (bandit) + dynamic taint tracking
Live watchdog kills rogue processes > 30 s CPU
Chaos‑tests inject tool failures; reward graceful degradation

7 Extending 🛠

New dataset – drop my.pkl into data/, flag --dataset my.
New metric – subclass scorer.BaseMetric, list in configs/default.yml.
New tool – add core/tools/foo.py exposing __call__(self, query).

8 Roadmap 🗺

☐ Hierarchical meta‑meta search
☐ GPU batch infer (Flash‑infer v3)
☐ Offline RL fine‑tune search policy with lineage replay

9 References 📚

S. Hu et al. “Automated Design of Agentic Systems” ICLR 2025
OpenAI “A Practical Guide to Building Agents” (2024)
Google ADK docs (2025)

View README on GitHub