Sparsi framework cuts AI agent token costs and latency using deterministic DAG tools

Developers building ReAct-based AI agents in production often face high token costs, slow latency, and occasional hallucinations due to LLMs re-deriving the same logic on every request. To address this, a team built Sparsi, a framework that offloads predictable sub-routines from the LLM prompt into deterministic Directed Acyclic Graph (DAG) structures called Macro-Tools. These DAGs can be exposed to existing agents as single Model Context Protocol tools, letting the agent delegate reliable, testable logic without repeated reasoning overhead. In a benchmark test on a five-step customer support triage task using 100 samples, Sparsi achieved 100% pipeline accuracy at 1.73 seconds average latency consuming 61,158 tokens, compared to 94% accuracy, 3.06 seconds, and 258,050 tokens for a pure LangChain ReAct agent. Sparsi is available as both a standalone pipeline solution and as a plug-in tool builder for existing agent architectures, with a pre-built operator library included.