Operama: Goal-Oriented Reliability and Self-Improvement for Multi-Agent Systems
Vishwanath Katharki (Operama), Sainyam Galhotra (Cornell University)
Engineering & Operations Evaluation & Benchmarking
Summary
A runtime reliability framework for multi-agent systems that decomposes goals into verifiable sub-goals, monitors execution, and automatically proposes policy updates without retraining.
Description
We present Operama, a runtime reliability and self-improvement framework for multi-agent AI systems. Modern agentic pipelines frequently degrade when task distributions change, tools fail, or prompts drift, yet existing frameworks provide limited support for diagnosing and correcting these failures. Operama introduces three core components: (1) a Goal Decomposition Engine (GDE) that translates high-level objectives into verifiable sub-goals, (2) a Runtime Reliability Monitor (RRM) that tracks agent execution and detects goal-level degradation, and (3) a Self-Improvement Loop (SIL) that analyzes execution traces and automatically proposes policy updates. In this demonstration we present a live end-to-end scenario where a multi-agent software engineering system plans, implements, tests, and deploys a microservice while Operama continuously monitors reliability and adapts the system when failures occur. Attendees will observe how Operama identifies failure points, generates new evaluation scenarios, and improves agent performance without retraining the underlying models.
