The crushing weight of a single tactical error in the 90th minute is a familiar tragedy for anyone who has played a football management simulation. For decades, the loop has remained the same: a human player analyzes data, clicks a menu, and hopes the simulation engine rewards their intuition. But a new experiment is shifting the decision-making power away from the human. The goal is no longer to see if a person can manage a team to glory, but to see if an autonomous AI agent can maintain the cognitive persistence required to run a professional club over an entire season.

The Architecture of an Autonomous League

Agentic FC arrives as an open-source football management simulation designed specifically as a playground for AI agency. With the release of version v0.1.0, the project has made its source code and executable binaries available via GitHub, ensuring that developers on macOS, Linux, and Windows can deploy the environment locally. Unlike traditional games, Agentic FC is not designed for human interaction through menus or controllers. Instead, it serves as an experimental environment where AI agents operate via the Model Context Protocol (MCP).

At its core, the simulation is powered by a local server written in Go. This server handles the heavy lifting of the game engine, while agents—such as those powered by Codex or Claude Code—connect to the server through the MCP interface. This protocol acts as the standardized bridge, allowing the AI to access external data and tools without needing a custom integration for every different model. Through MCP, the agent performs the duties of a manager: it reads the current league standings, analyzes the specific attributes of its players, and evaluates the overall flow of the season. Once the data is ingested, the agent issues concrete instructions to set tactics and manage the club's operations independently.

While the AI handles the management, the human user occupies the role of an observer. This is achieved through a Text User Interface (TUI) console. Rather than relying on high-fidelity graphics, the TUI provides a stream of textual data that mirrors the experience of a real-world sports observer. Users can read simulated media articles, track the league table, review player rosters, and follow real-time match commentary. The TUI also includes a replay-style observation screen, allowing humans to audit the decisions the agent made and see how those choices translated into results on the pitch.

Solving the Persistence Problem in AI Agency

For most LLM-based applications, the primary constraint is the cost and complexity of maintaining a consistent state over a long period. Most AI interactions are ephemeral—a prompt, a response, and a reset. Agentic FC attempts to break this cycle by physically separating the execution environment from the AI's reasoning process. By using a Go-based server to hold the world state, the AI does not need to remember every detail of the league in its own context window; it simply queries the server via MCP whenever it needs a factual update.

This architectural choice allows the simulation to implement deep, interlocking systems that would typically confuse a standard chatbot. The environment includes a fully realized ecosystem of league matches, cup tournaments, player transfers, and contract negotiations. It even simulates the volatility of professional sports, incorporating player injuries and a board-of-directors trust system that can end a manager's career. To ensure these experiments are reproducible, the system utilizes seeded worlds. By using a specific seed value, the simulation creates a consistent virtual universe that can be saved and restarted, allowing developers to test how different AI agents handle the exact same set of circumstances.

This shift reveals a fundamental change in how we measure AI capability. In a standard benchmark, an AI is judged on its ability to provide the correct answer to a static question. In Agentic FC, the benchmark is persistence. The agent's success is not determined by a single clever response, but by its ability to manage a complex web of variables—player morale, financial constraints, and tactical shifts—over hundreds of simulated days. The tension moves from the quality of the output to the stability of the agency.

The integration of a Go server with the MCP interface proves that AI agents can move beyond the chat box and into the role of operational entities. By managing a simulated football club, these agents demonstrate that the next frontier of AI is not better conversation, but the ability to maintain a long-term objective within a volatile, stateful environment.