- Published on
Why harnesses matter for energy (and token) efficiency
- Authors
- Name
- DeepHorus
Agent harnesses can significantly affect inference energy and token efficiency, according to a month-long analysis by cloud provider Neuralwatt. The company found that clients cache input differently, with uncached input ranging from 1.6% to 11.1%. At 92% cached, 8 of every 100 input tokens are uncached, versus 3 at 97%, creating about 2.7 times the prefill work.
Neuralwattâs traffic is overwhelmingly agentic. A typical coding-agent turn sends 100,000 tokens of conversation and receives only a few hundred tokens back. Because prompt caching requires an exact prefix, a harness that edits already-sent text forces the server to recompute from the first difference onward, and a change near the beginningâeven a timestampâcan be costly.
The study covered 30 days from September 8 to October 7, 2026, and hundreds of clients, including OpenCode, Oh My Pi, Pi, OpenAI SDKs, and custom apps. OpenCode alone serves about one in four customers. Deepseek-harness broke cache on about 1 turn in 100, while OpenCode and Pi broke it on about 3 in 100.
OpenCode 2.x looked better aloneâ2.2% uncached versus 3.0%âbut mostly because of model mix; on the same model, both versions cache about equally well. Energy costs are direct. A cache miss roughly doubles a requestâs energy, and on GLM-5.3 it is about 2.5 times as much.
Deep sessions amplify the problem: past 128,000 tokens, one turn that re-reads more than half its prompt costs as much prefill as about 600 normal turns at that depth. Most cache breaks are fixable. Every break in Neuralwattâs lab tests came from a harness editing what it had already sent.
OpenCode 2.x fixes most 1.x breaks, though the difference is not visible in production. Harness design deserves closer attention alongside model and hardware choices. Better prompt construction can cut GPU work and energy use without changing the model.
