LUAF — Large-scale Unified Agent Foundry

LUAF turns a business brief into a research → build → validate → launch pipeline for autonomous business units. You describe a use case (or leave it blank for an AI-generated idea); LUAF runs web search, optional planner/designer LLM steps, validation (with auto pip-install on missing imports), and publish to swarms.world with optional Solana tokenization.

Repository: github.com/Euroswarms-Institute/LUAF

---

What it does

• Topic / brief: You provide a short description of the autonomous unit you want (e.g. "DeFi backtester for ETH/USDT") or press Enter to have the LLM generate a monetizable idea.
• Research: DuckDuckGo (and optional multi-hop RAG) gathers context.
• Plan (optional): Planner produces a structured plan; executor fills a template.
• Design: Designer LLM produces a full Python agent (name, ticker, code, requirements, use cases). Quality packages (e.g. swarms, loguru, tinydb, requests) are enforced per topic.
• Validate: Generated script is run in a subprocess. On ModuleNotFoundError, LUAF runs pip install <module> and retries. Optional urllib3/Retry compatibility fix is applied. If validation fails, the user can choose to publish anyway or validate manually.
• Publish: Payload is sent to POST /api/add-agent with Bearer token (SWARMS_API_KEY). Dry-run by default; set LUAF_DRY_RUN=0 and provide SOLANA_PRIVATE_KEY for tokenized launch.
• Persistent mode: Loop until wallet balance reaches a target SOL; each iteration can use a new or fixed topic (env/file list).

---

Environment variables (summary)

See repo README and in-code comments for the full list.

---

Usage tutorials

1. Interactive menu (TUI or CLI)

BASH
python LUAF.py  

• If Textual is installed: TUI with balance, target SOL, current brief, live log. Use s to stop, q to quit.
• Otherwise: text menu — 1 = Pipeline, 2 = Persistent, 0 = Exit.

2. Single pipeline run (no TUI)

BASH
python LUAF.py --no-tui -o  

• Prompts: business use case or brief (Enter = generate), unit name (Enter = auto), ticker (Enter = auto).
• Runs one designer cycle (up to MAX_STEPS with validation retries). Saves generated unit to <Name>.py or generated_agent_step<N>.py. On validation failure, asks whether to publish without validation and mentions manual validation.

3. Persistent (autonomous loop)

BASH
python LUAF.py --persistent  
# or  
LUAF_MODE=persistent python LUAF.py  

• Runs until Solana balance >= LUAF_PERSISTENT_TARGET_SOL (default 10) or user stops.
• Each iteration: get next topic (single / env list / file), search, designer, validate, publish if funds, run unit once, optional delayed claim.

4. Self-train pipeline

BASH
python LUAF.py --self-train "Your topic here"  
python LUAF.py --self-train   # uses LUAF_TOPIC or default  

5. Legacy positional

BASH
python LUAF.py run        # same as --once  
python LUAF.py persistent # same as --persistent  

6. Help

BASH
python LUAF.py --help  

Shows description, epilog examples, and all flags (--no-tui, --once, --persistent, --self-train).

---

Requirements (for running LUAF)

Core: requests, python-dotenv, loguru, ddgs. Optional: textual (TUI), sentence-transformers (RAG), swarms, toolbox, planner, executor, tinydb (designer quality-packages). See requirements.txt in the repo.