PRISM: Optimizing Scalable Agentic Workflows

PRISM is built for the future of scalable AI agents. As agents become part of software development, they will need more than access to bigger models and larger tool registries. They will need infrastructure that measures and optimizes how they reason, choose tools, recover from failures, and execute workflows. PRISM applies retrieval to the agent tool layer. Instead of loading every tool schema into the agent prompt, PRISM retrieves a focused set of tools for each task. The demo compares a baseline all-tools agent against the PRISM agent on a multi-step workflow involving search, scraping, analysis, and email generation. The result is a measurable reduction in tools loaded, schema tokens, duplicate calls, category mismatches, confusion score, and latency. PRISM also demonstrates failure recovery by falling back from a failed scraper to an HTML extractor. The project is deployed publicly on Hugging Face Spaces and benchmarked on AMD Developer Cloud using an AMD Instinct MI300X GPU with ROCm/HIP-enabled PyTorch. It is positioned as a compiler-like optimization layer for agentic software development, helping future coding agents scale efficiently as their available tools grow.