Agent AI Costs Are Determined by the "Harness" (Control Layer) — Results of a Large-Scale Comparative Experiment
A research paper is drawing attention for showing that in enterprise agent AI development, the design of the orchestration layer (the "harness") determines cost efficiency more than the model itself. "The Harness Effect," published by a team at Writer, conducted a controlled experiment using 22 evaluation tasks and 6 foundation models (Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, and Palmyra X6), swapping out only the orchestration layer.
The results were striking. With an optimized harness, cost per task dropped 41% ($0.21 → $0.12), median execution time fell 44% (48 seconds → 27 seconds), and token consumption decreased 38% (14.2k → 8.8k) — all while maintaining equal or better task completion quality (0.78 → 0.81).
What's particularly interesting is that this efficiency gain is model-invariant: cost reductions of 33–61% were observed across all models. Meanwhile, the degree of quality improvement correlates with the model's underlying capability — the stronger the baseline, the greater the benefit from the harness. The team named this phenomenon "harness leverage," and it makes a case for investing as much effort in control layer design as in model selection.
As token prices continue to fall, approaches that rely on "token maxing" (burning huge amounts of tokens to buy capability) drive up total spending. The paper argues that improving the orchestration layer is the one efficiency lever that multiplies across every model.
NVIDIA and Microsoft Co-Develop a Desk-side AI Supercomputer for Windows
NVIDIA has jointly announced the "NVIDIA DGX Station for Windows," a desk-side AI supercomputer for developing and running AI agents in Windows environments, together with Microsoft. Powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, it is said to be capable of running state-of-the-art AI models with up to 1 trillion parameters locally.
Its biggest feature is that enterprises can run and develop "hundreds of AI agents" in isolation within their own Windows environments. Being able to perform large-scale AI inference without relying on the cloud, while keeping sensitive data on-premises, is a major advantage for organizations with strict security and compliance requirements.
Desk-side AI development machines have mostly run Linux until now, but native Windows support could make building AI agents far more accessible to enterprise developers.
AI Finds 7 Real Bugs Lurking in Cloudflare's Cryptography Library
A case reported on Qiita is drawing attention: AI managed to find seven real bugs in CIRCL, Cloudflare's open-source cryptography library.
Among the bugs found were logic errors easy for human eyes to miss, such as misuse of Go's bitwise OR operator (`|`). The case handling in a selector expression for `modeBase | modeAuth` and `modePSK | modeAuthPSK` was incomplete, allowing certain mode combinations to enter the wrong branch.
Bugs in cryptography libraries are extremely serious from a security standpoint, but their code complexity and deep mathematical background can make them difficult to catch through human review alone. That AI-powered code auditing works at a practical level raises expectations for it as a new tool in security engineering.
Nature Reports — "Humanizer" Tools That Erase Traces of AI-Generated Text Emerge
According to a report in the scientific journal Nature, tools called "Humanizers" that strip away the "AI-ness" of AI-written text have emerged, sparking debate in academia.
As AI detection tools improve in accuracy, the tools designed to evade them keep evolving as well, deepening the cat-and-mouse game. Educational institutions in particular are concerned about students using Humanizers to slip past detection when submitting assignments written with AI.
The problem runs deep. There is a fundamental trade-off between detection accuracy for AI-generated text and the diversity of human writing styles, making perfect detection theoretically impossible. In education and publishing, discussion has begun on shifting away from detection-dependent approaches toward evaluating the writing process itself.
China Eliminates 12,200 University Programs, Replacing Many with AI Degrees
According to Forbes, China has carried out a sweeping education reform, cutting 12,200 university programs nationwide and replacing many of them with AI-related degree programs.
This move clearly signals that the Chinese government has made talent development in AI a top priority. The figure of 12,200 represents a structural transformation that goes beyond mere curriculum adjustment: the country is building a system to mass-produce AI talent in exchange for shrinking traditional academic fields.
At the same time, there are concerns both at home and abroad about how such a rapid transition will affect educational quality and academic diversity. Still, there is no doubt that talent supply is the key in national AI strategy, and China's move could influence education policy in other countries as well.
MTPLX V2 — Running Qwen 3.6 27B at 82 TPS on a MacBook Pro
A fast inference runtime for MLX models, "MTPLX V2," has been announced on Reddit's LocalLLaMA community. It is claimed to run Qwen 3.6 27B at 82 TPS (tokens per second) on a MacBook Pro M5 Max.
The biggest improvement in V2, released just one month after V1, is "Turbo Mode": by combining a verification-specialized quantized matmul kernel with precompiled verification steps, it achieves 82 TPS at temperature 0.6. It also brings an improved SSD KV cache and better long-context tool calling.
The practicality of local LLMs is directly tied to inference speed. Breaking 80 TPS with a 27B-class model puts everyday text generation and coding assistance at a level locally that is on par with cloud APIs. It's a sign that AI inference optimization on Apple Silicon continues to advance.
---