Routing and Governing AI Coding Agents to Control Enterprise Costs

Summary

AI coding-agent harnesses determine which model responds, what context it can read, how prompt caches are used, and which subagents run, making them a major driver of enterprise model costs. This paper argues that enterprises using vendor-built harnesses or untuned internal defaults inherit those routing and pricing decisions. It presents Jev, a customizable router with calibrated probabilities that classifies each prompt according to an enterprise-defined taxonomy of agentic requests. Because a user turn may contain multiple requests tied to one model’s prompt cache, Jev moves work only at points where a conversation does not need to rebuild that cache: session start, side lanes, and subagent launch. The authors derive a price-based crossover showing that, in long, tool-heavy sessions, a higher-priced model can cost less than the next tier when repricing is considered. Repricing about 10,000 real sessions from public datasets supports this result. In an emulated enterprise with 10,000 seats, the router recovers 14% to 21% of model spend at the listed prices used in the study, equivalent to $3.3 million to $5.0 million annually. The paper also examines risks across 20 harnesses, quantifies dependence on one vendor’s models, and proposes an internal control plane plus a staged path toward deciding whether to own the harness.