Antrocon — Multi-Agent LLM Orchestration System
Overview Antrocon is a multi-agent orchestration system for large language models (LLMs), designed for research into autonomous agent collaboration, structured dialog management, and long-term memory retrieval. It coordinates two autonomous LLM agents — a Builder (Agent A) and a Critic (Agent B) — in a turn-based dialog loop, while a human Operator (the "Papa Agent") monitors and steers the system through a real-time web dashboard. Architecture The system runs locally on a Windows host with two instances of llama.cpp serving quantised open-weight models via an OpenAI-compatible API. A central Python orchestrator manages dialog flow, tool exec
Resource information
- Format
- zip
- Original repository
- 18648837
- DOI
- 10.5281/zenodo.18648837
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- mit-license
- Last record update
- Nov 5, 2026
About this resource
Overview Antrocon is a multi-agent orchestration system for large language models (LLMs), designed for research into autonomous agent collaboration, structured dialog management, and long-term memory retrieval. It coordinates two autonomous LLM agents — a Builder (Agent A) and a Critic (Agent B) — in a turn-based dialog loop, while a human Operator (the "Papa Agent") monitors and steers the system through a real-time web dashboard. Architecture The system runs locally on a Windows host with two instances of llama.cpp serving quantised open-weight models via an OpenAI-compatible API. A central Python orchestrator manages dialog flow, tool execution, context compression, and persistent state. An isolated Ubuntu VM provides a sandboxed execution environment for shell commands and browser automation. Core Capabilities Multi-agent dialog loop — Builder proposes actions; Critic reviews and suggests corrections. Both agents share a common tool set (shell execution, browser control, file I/O) and can invoke tools autonomously. Context compression — An iterative compression engine summarises dialog history into structured "continuation guides" when the context window limit is approached, preserving task continuity across long sessions. Long-term vector memory (LTM) — A concurrency-safe, file-locked vector database (sentence-transformers + NumPy) stores and retrieves relevant prior knowledge across dialog sessions. Explorative testing framework — 126 automated scenarios validate API endpoints, authentication, role-based access control, rate limiting, and error handling. Web dashboard — A FastAPI backend with a single-page application provides live dialog monitoring, seed management, LTM inspection, and explorative test execution. Tree/Branch/Seed (TBS) planner — A hierarchical task decomposition model that allows agents to negotiate multi-step plans and execute them as independent dialog sessions. Technical Stack Python 3.10+ (orchestrator, dashboard, CLI) llama.cpp with Qwen3-8B (quantised, local inference) FastAPI & Uvicorn (web dashboard) sentence-transformers (LTM embeddings) Playwright (browser automation in sandbox VM) SSH/SFTP (host-to-VM communication) Scope and Limitations Antrocon is a research and demonstration system. It is not intended for production deployment or safety-critical applications. The system has been developed and tested in a single-user, local-inference setting. Its primary purpose is to explore architectural patterns for multi-agent LLM orchestration, including context management under finite context windows, persistent memory across sessions, and structured human oversight of autonomous agents. The codebase is documented bilingually (German and English) and includes comprehensive governance documentation covering GDPR compliance, security policies, and the EU AI Act. License Source code: MIT License. Documentation: Creative Commons Attribution 4.0 International (CC BY 4.0).