Agent harness affects cache efficiency and energy use

Neuralwatt's analysis of a month of cloud traffic found agent harnesses caused 1.6%–11.1% uncached input. A cache miss roughly doubles request energy, 2.5x on GLM-5.3; in deep sessions beyond 128k tokens, re-reading over half the prompt costs as much prefill as ~600 normal turns. All cache breaks came from harnesses editing sent content, so most issues are fixable.

• Open Models Debate: Weighing Safety Against Democratization2026-10-13
• Prime Agent rewritten in Rust by 2,000-agent swarm2026-10-10
• CivBench benchmarks LLMs playing Civilization V2026-10-06
• GLM 5.3 launches on Amazon Bedrock2026-10-06
• Anthropic warns GLM-5.3 may spread advanced cyber capabilities2026-09-29