Medusa
Reusable agent runtime · AIBC
AIBC's reusable agentic engine. Security gateway in front of every tool call, MCP-native transport, and a hybrid neuro-symbolic cascade that combines cheap heuristics with multi-LLM validation before any irreversible action. Powers DACv2.
Problem
Every agent product at AIBC needed the same underlying machinery — tool access, LLM routing, safety checks — and rebuilding it per product meant re-solving the same trust problems each time. The dangerous part of an agent isn't the chat; it's the tool call that deletes, sends, or pays.
Approach
Medusa is the machinery extracted into one runtime. A security gateway sits in front of every tool call — nothing reaches a tool without passing it. Transport is MCP-native, with agent-to-agent (A2A) messaging so agents can delegate to each other.
LLM adapters cover four providers (Anthropic, OpenAI, Gemini, Ollama), so an agent can be tuned for cost, quality, or run fully local. Before any irreversible action, a hybrid neuro-symbolic cascade runs: cheap symbolic heuristics first, multi-LLM validation only when the heuristics aren't sure — with human-in-the-loop gates as the final layer.
Runtime is Node.js/TypeScript with an Express REST API and Server-Sent Events for live agent streaming, plus an operator dashboard for oversight.
Outcomes
- Powers DACv2, AIBC's autonomous multi-agent platform.
- First stress-test task built on top: an Unsubscribing Agent used to probe the limits of safe autonomy.
- Open source on GitHub.