Paper Prices Enterprise Cost of Locking Coding Agents to One Harness

A paper posted to arXiv on September 24, 2026 by Arian Abbasi, Alan Aqrawi and Ted Kwartler takes aim at a detail enterprises buying AI coding agents tend to overlook: the harness, the software wrapper that actually runs the agent, decides more of the bill than the underlying model price. Harnesses choose which model answers a given prompt, what context it reads, how the prompt cache gets reused and which subagents fire, so the harness effectively sets the rate on the price sheet and the volume bought at that rate. Most enterprises do not build their own harness. They buy one, commonly Anthropic’s Claude Code or OpenAI’s Codex, and if they leave it at its defaults they inherit every routing decision the vendor made, along with the invoice that follows.

The authors’ response is a router built around a classifier they call Jev, which assigns calibrated probabilities to every incoming prompt against a taxonomy of coding-agent request types that a buyer can define. Because a single user turn can generate many requests running over one model’s prompt cache, the router only shifts work at three points, session start, side lanes, and subagent launch, where no active conversation would need its cache rebuilt from scratch. The paper also works out, from published price sheets, when switching models mid-task actually pays off, and finds a crossover: on sessions that run long and lean heavily on tools, the most expensive model on the price list can end up cheaper than the next tier down, once cache costs are counted in.

The numbers

Testing the approach against roughly 10,000 real sessions drawn from public datasets, the authors confirm that crossover holds in practice. Applied to an emulated enterprise of 10,000 seats, with user behavior drawn from the same datasets, the router recovers 14 to 21% of model spend at Anthropic’s list prices as of September 21, 2026, which the paper translates to $3.3 million to $5.0 million a year for that size of deployment.

Beyond the router itself, the paper maps risk across twenty different harnesses and prices out what it costs an enterprise to depend on a single vendor’s models. It closes with a control plane enterprises can run themselves starting now, plus a staged plan for deciding later whether to build and own a harness outright rather than keep renting one.