Tool-use agents are expensive in a specific way: at every hop, the language model re-reads the current state and spends seconds thinking, even though the set of possible actions is finite and many decisions repeat. This per-decision cost is the problem addressed by NIOCIS, a routing layer described in a new arXiv preprint.
NIOCIS works by folding execution logs into the routing process. Rather than letting the model start from scratch each time, the routing layer uses accumulated logs to inform next-hop decisions. The abstract suggests this can reduce the need for repeated re-reading and thinking, though the full mechanism is not detailed in the available text.
Because the source is a single preprint abstract, independent validation is not yet available. The contribution is positioned as a systems-level fix: a routing layer that sits between the agent and its tools, making recurring decisions cheaper without changing the underlying model.