Specialized multi-agent autonomous coordination for complex project execution using balanced cooperation, dynamic virtualized playgrounds, and extensible tool sets
Αυτόματος συντονισμός πολλαπλών εξειδικευμένων πρακτόρων για την εκτέλεση πολύπλοκων έργων με χρήση ισορροπημένης συνεργασίας, δυναμικών εικονικών μηχανών και επεκτάσιμων συνόλων εργαλείων
View/
Keywords
Multi-agent systems ; LLM orchestration ; Model Context Protocol ; Virtualized sandboxing ; OpenNebula ; Token optimization ; Autonomous AIAbstract
The rapid advancement of Large Language Models (LLMs) has, at the same time, initiated a shift from simple
conversational interfaces to autonomous Multi-Agent Systems (MAS) capable of executing complex technical
workflows. However, current orchestration frameworks face three critical bottlenecks: the "memory wall" where
context degradation leads to hallucinations in long-horizon tasks; vendor lock-in through tight coupling with
proprietary APIs; and significant security risks associated with running agent-generated code without
hardware-level isolation.
This thesis introduces SandBot (an Agent Sandboxing & Orchestration Framework), an enterprise-grade MAS
architecture designed to bridge the gap between experimental research and secure production deployment.
SandBot addresses context exhaustion through a "Six-Section Prompt Architecture" and an algorithmic "Tool
Description Deduplication" technique that reduces Model Context Protocol (MCP) overhead by approximately
47%. To ensure operational safety, the framework programmatically integrates with OpenNebula to provision
temporary, network-isolated MicroVMs for agentic tasks, enforcing isolation that mitigates Arbitrary Code
Execution (ACE) risks.
Central to the framework is an Orchestrator-Delegate pattern characterized by balanced peer-to-peer
cooperation. This topology replaces rigid hierarchies with a flat network of specialized peers - standardized as
"Atomic Agents" - that interact via isolated namespaces. By decoupling orchestration logic from the underlying
cognitive models (Bring-Your-Own-Model), SandBot provides a sovereign, future-proof, and secure path for the
autonomous execution of complex technical projects.