Rules Engines Are Back, and Your Agent Is the Interface, Not the Executor
UiPath, Tuesday, November 10th, 2026
UiPath argues multi-step AI agents are safer when an LLM proposes actions, a DMN rules engine approves them and an orchestrator executes.
UiPath notes that multi-step agents fail in production when one bad call cascades through later steps; an agent with 85% per-step success has roughly a 23% chance of completing a nine-step workflow correctly.
Function calling and schema-constrained structured outputs catch malformed requests but not wrong business decisions.
The proposed pattern splits the work: the LLM acts as the interface and produces a structured request, a rules engine authored in Decision Model and Notation approves it, routes it to a human or rejects it, and a durable orchestrator handles execution, retries, approvals, audit trail and recovery.
The post says this is unnecessary for simple two-step workflows and targets long-running agents on real backends.