Archived -- This project is no longer actively maintained.
Built March 2024. An early experiment in multi-agent coordination using local LLM inference.
AAS is a multi-agent autonomous swarm system designed around hierarchical task decomposition and shared project state. Agents are organized in a tree structure where each agent has a role, mission, and set of permitted functions. They coordinate through a Unified Project Framework (UPF) -- a shared workspace of typed documents that agents can read and write based on access permissions.
Core concepts:
- Agent hierarchy -- Agents can create and manage subordinate agents, forming a tree of specialized workers
- Function-based tool use -- Each agent has an explicit list of functions it can call, defined as name + usage instructions
- Unified Project Framework -- Shared typed object store (text documents, etc.) that agents collaborate through
- Evaluation Oversight Mechanism (EOM) -- Planned evaluation layer for monitoring agent outputs
- LLM Integration (LMIS) -- Local inference via LLamaSharp with CUDA acceleration
- C# / .NET 8
- LLamaSharp 0.10.0 -- local LLM inference
- CUDA 12 backend
AgentSystems/ Agent base class, function definitions, hierarchy
AgentFunctions/ Tool/function scaffolding for agents
Evaluation(EOM)/ Evaluation and oversight (scaffolded)
LlmIntegration(LMIS)/ Local LLM inference layer (scaffolded)
UnifiedProject(UPF)/ Shared project state and typed documents
This was an early prototype exploring multi-agent coordination patterns -- hierarchical delegation, shared state, and function-based tool use. The architecture is defined but several subsystems (EOM, LMIS, agent functions) remain scaffolded. It served as a stepping stone toward later agent infrastructure work.
MIT