Researchers from Renmin University have published an extensive 149-page review dedicated to long-term agents. The work proposes two main development trajectories for next-generation artificial intelligence agents: Harness Engineering for improving runtime capabilities and Model Optimization, which gradually integrates execution capabilities back into the model itself through training.
This comprehensive review, prepared by the GAIR Research Center at Renmin University in collaboration with Peking University, Tsinghua University, Sun Yat-sen University, The Hong Kong University of Science and Technology, and the National University of Singapore, presents the first systematic framework for understanding this evolving field. A key characteristic defining these agents is their ability to organize interdependent decisions into coherent trajectories.
The main contribution of the review is the presentation of a three-tiered hierarchy of task complexity and corresponding capability hierarchy. H1-level tasks, solved within a single context window, require maintaining planning-action-feedback-correction cycles and correspond to C1 interactive reasoning capability. H2-level tasks, spanning hours or days, necessitate history compression, progress retention, and inter-session transfer, exceeding the capabilities of extended context windows; here, 'Context Rot' is identified as the fundamental bottleneck, corresponding to C2 state and memory. The most complex tasks, H3, represent task flows across different processes, where assignments arrive continuously as the environment, tools, and goals change, requiring the accumulation of experience into reusable skills and continuous refinement, corresponding to C3 experience accumulation and continuous learning.
The review provides quantitative evidence of capability advancement using public METR data. Analysis shows that the time taken to complete 50% of tasks, measured in equivalent human expert time for tasks the agent completes with approximately 50% success, doubles every 196.5 days (about 7 months) across the entire dataset. Focusing on the period after 2023, this accelerates to approximately 130.8 days (4 months). As of 2026, GPT-3 demonstrated about 9 seconds of task execution time, GPT-4—about 4 minutes, o3—about 2 hours, and Claude Opus 4.6—about 12 hours. This transition from minutes to hours and tens of hours reflects the escalation of capabilities from H1 to H2 and further to H3.
The authors describe three eras of AI agent development. The first, the prompt engineering era (2020–2023), relied on chain-of-thought and task decomposition within a single model call. The next, the context engineering era (2023–2025), expanded control over what the model could see through RAG, function calling, and memory. Finally, the era of Harness Engineering arrives, where the boundaries of capabilities are defined by the system architecture surrounding the model. The authors argue that the key insight is that the domain of controlled engineering solutions is gradually expanding outwards, and long-term capabilities depend less on improvements in the model's intelligence and more on the sophistication of the surrounding system architecture. This 149-page review establishes a unified terminology and conceptual basis for this rapidly growing but previously unsystematized field, providing researchers and practitioners with a roadmap for the next stage of agent capability development.
Holmirzaev Firuz Suyun's son's dissertation on the topic 'Development of technology for obtaining fast-drying organic coatings for road signs' will be defended.
This dissertation was written with the aim of obtaining the academic degree of Doctor of Philosophy (PhD) in the specialty 02.00.14 - Technology of organic substances and materials based on them (technical sciences). The defense will take place at the meeting of the Scientific Council for awarding academic degrees DSc.30/2025.27.12.K/T.01.01 held at the Tashkent Chemical Technology Research Institute on July 30, 2026, at 11:00 AM.
The defense will take place at the Tashkent Chemical Technology Research Institute in Shurobozor, Ibrat MFI, Tashkent district. For contact, the phone number is (90) 964 96 05, and email addresses are ooo_tniixt@mail.ru and TKTITI@exat.uz.
Chengdu CAS 4.0 Technology has supplied over a thousand radio frequency (RF) testing systems since 2018, becoming the first Chinese company whose RF testers possess full intellectual property for integration into chip-scale production lines. This achievement concludes an eight-year period of market development, where domestic brands held less than a 10% share.
Founder and CEO Huang Xuguang identified a development opportunity while working on microwave components. He noted that almost all microwave test equipment was imported, characterized by high costs, long delivery times, and slow adaptation to local industry needs. The first three years were dedicated entirely to research and development without active market entry.
The team systematically studied the specifications of leading international products in high-density multichannel RF equipment design and calibration algorithms. The initial phase focused solely on addressing gaps caused by the lack of domestic alternatives. The second phase evolved into jointly defining requirements with Chinese industrial clients, enabling the creation of unique products tailored to local manufacturing needs rather than simply copying foreign solutions.
The co-definition model with local chip manufacturers ensures customized tools and local support unavailable from foreign competitors. Three months of on-site integration led to certification from Apple and Tesla, and the equipment is now used on iPhone and Tesla production lines.
The product range includes four categories of instruments: signal generators, signal analyzers, vector network analyzers, and vector transmitters, which utilize a modular USB and PXIe architecture. CAS 4.0 offers the most comprehensive set of products and the widest frequency coverage among domestic brands with a PXI architecture. The self-developed board-based RF chip ATE testing system can perform mass testing of filters, switches, and power amplifier modules in a single setup, significantly increasing testing efficiency and reducing costs compared to traditional methods.
Key technical parameters meet world-class standards: the WiFi-7 chip's EVM coefficient reaches below -48 dB, and the PXIe vector network analyzer demonstrates a dynamic range of 132 dB with a scanning speed that significantly surpasses traditional benchtop instruments. Sixty percent of the company's staff is engaged in R&D, and annual investments exceed 20 million yuan. Plans include expanding up to 40 GHz by the end of 2026 for LEO satellites, and in the medium term, up to 80 GHz for 6G. The next three to five years are defined as a golden period for replacing RF testing equipment with domestic alternatives, with AI integration throughout the lifecycle further reducing costs.
This website employs a specialized security service designed to protect against malicious bots. This page is displayed during the verification process to ensure that the user is a human and not an automated bot.
The security service is used to prevent attacks from malicious bot programs. The display of this page is part of a process that confirms that the site visitor is not a bot.
