Post by Harish Sundarakumar

Incoming SWE Intern @ HPE Juniper WAN | AI Infrastructure & Distributed Systems | MCS @ UC Irvine

Most multi-agent systems don't struggle because the agents are bad. They struggle because every new capability introduces new coordination logic, dependencies, and integrations. During our UC Irvine MCS capstone, in partnership with Accenture, our team set out to solve a systems problem: How do you coordinate independently deployed AI services without constantly rewriting orchestration logic? We built the Universal Utility Orchestrator (UUO), a cloud-native orchestration layer that dynamically discovers services through an agent registry and constructs execution plans at runtime. Rather than tightly coupling orchestration logic to agent implementations, we designed the system around runtime discovery, dependency-aware scheduling, and loosely coupled service integration. Some of the engineering challenges we tackled: šŸ”¹ Dynamic agent discovery — Agents can be hosted anywhere and registered through metadata rather than code changes. New capabilities can be onboarded without redeploying the orchestrator. šŸ”¹ Dependency-aware execution — We model workflows as DAGs and use Kahn's algorithm to detect cycles and generate topologically ordered execution layers for parallel execution. šŸ”¹ Reactive and proactive workflows — The same orchestration engine supports both user-driven requests and autonomous background triggers. šŸ”¹ Observability and auditability — Execution traces, plans, and conversation history are persisted to Cosmos DB for end-to-end visibility. šŸ”¹ Lightweight cloud-native runtime — Built with Azure Functions, OpenAI SDK, async HTTP execution via httpx, Pydantic validation, and managed identities rather than heavyweight orchestration frameworks. One of the biggest lessons from the project was that orchestration becomes the bottleneck long before individual agents do. Building reliable scheduling, dependency management, observability, and service coordination turned out to be significantly harder than building the agents themselves. Huge thanks to our Accenture mentors and sponsors Shawna Tuli, Mo Nomeli PhD, and Vishrut Chokshi for their guidance throughout the project, and to my teammates Sanket Landge, Arya Gupta, and Shaun Lim for making this project possible. The core infrastructure is now open source. I'll share the repository in the comments. #SoftwareEngineering #DistributedSystems #SystemArchitecture #BackendEngineering #CloudNative #Azure #AIInfrastructure #OpenSource

Post contentPost contentPost content