1. Introduction: The Future of Work Automation - LLM Messenger
In the midst of endless data pouring in daily and complex business processes, modern office workers are always hungry for efficient management tools. While automation in the past was merely simple macros that moved according to fixed rules, we are now entering the era of 'autonomous agents' where AI makes its own judgments, exchanges tools, and completes execution.
After several months of hard work, I created a complete agent management program called 'Troy'. It performs the role of an intelligent business partner that receives user commands, handles complex pipeline editing, and optimizes business workflows on its own.
2. Six Family Members Under One Roof - Six LLMs Running Locally
The most distinctive feature of the Troy system is that it does not rely on massive cloud servers or expensive subscriptions. Instead, six different language models (SLM, which developers call LLM) operate simultaneously within the user's local network. Users can freely choose and use optimized models such as Fast, Quality, Deep, Minai, Silicon, and Flower depending on their purpose.
This innovation is possible because we built an AI environment within the local network by independently utilizing the 'Llama.cpp' server. Through clustering, it guarantees powerful performance without external internet connection, and achieves two goals at once: security without worrying about data leakage and fast response speed.
3. The Self-Evolving 'Laboratory' - An Agent that Performs AB Testing
While traditional software has fixed functionality once a developer writes the code, Troy's agents evolve on their own. They discover new skills from GitHub and similar platforms, and constantly perform 'AB testing' comparing existing features and performance. The model and code that win here are then applied to the next task.
The core of this process lies in 'Code DNA Matching' and 'Code Fingerprinting' technology. When an agent performs a specific task (such as content extraction), it automatically judges whether the existing 'Firecrawl' method is efficient or whether the 'Headless Browser' method is better, and makes replacements accordingly. Without developer intervention, the system as a 'laboratory' continues to operate and perfect the optimal 'recipe'. This means that even after a developer leaves, the software can maintain its own vitality and stay up-to-date.
4. Just Provide an Excel File or API Document and a Website Appears? The Agent's Alchemy
Troy's execution power goes beyond simple text generation or code writing, reaching into the realm of 'infrastructure construction'. AI has reached the level where, given an Excel file or API as input, it can perform reverse analysis and completely reconstruct a functioning website, designing APIs and databases.
As a real example, it clones Naver Works, generates a Swagger API-based site, and perfectly sets up everything from database structure to actual service infrastructure. This transcends simple 'No-code' tools and symbolizes a shift to an 'Agent-built' era where agents directly design and construct services. It provides disruptive productivity that allows non-professionals to easily build complex system infrastructure.
5. The Birth of 'P2P ERP' Without Servers
The most innovative point is the 'P2P (Peer-to-Peer) style ERP' system that has no centralized server. Troy operates business processes and manages warehouse inventory by directly exchanging data between agents without a server.
This approach not only reduces monthly cloud server costs for companies to nearly zero, but also fundamentally solves the limitations of the existing security model where all data is exposed if the central server is breached. The most economical and safest business operation model is implemented through AI in your pocket.
6. Workflow Recording, Remote Control, and the Magic of MTLS
Troy learns user work patterns by recording the execution process of specific functions and saving them as workflows, providing orchestration capabilities to capture data from remote computers or directly control them through remote access.
Particularly noteworthy is the 'MTLS Mini Server' feature. This immediately matches functions being developed in the local environment to specific domains and connects them as actual services. Without complex server configuration or deployment processes, this feature allows local AI agents to immediately become public services communicating with the outside world, dramatically shortening the time it takes to apply ideas to practical environments.
7. Why Modify an Agent That Doesn't Suit Me When I Can Write One from Scratch
AI is no longer a simple tool that we call only when needed. It has evolved into an autonomous 'agent' that learns new technologies on its own, exchanges optimal tools to fit the situation, and operates entire businesses without servers.
All programs are divided into CLI fragments and configured to be used by AI only through wiki and git.
It works well on Windows, Mac, and Ubuntu.
