Core Viewpoint - The emergence of "long-horizon agents" is reshaping the software engineering paradigm, moving from deterministic code-based systems to models that operate as black boxes, requiring real-time execution to understand their behavior [2][3][6]. Group 1: Long-Horizon Agents - Long-horizon agents are seen as a turning point in AI, with predictions that their adoption will accelerate by the end of 2025 to 2026 [2]. - These agents function more like "digital employees," capable of executing tasks over extended periods, learning from trial and error, and self-correcting [2][3]. - The transition to long-horizon agents may challenge traditional software companies, similar to the shift from on-premises to cloud solutions, where not all companies successfully adapted [2][3]. Group 2: Differences in Software Development - Traditional software development relies on deterministic logic written in code, while agent-based systems introduce non-deterministic behavior, making it necessary to observe their real-time execution to understand their operations [30][32]. - The concept of "tracing" has become crucial in agent systems, allowing developers to track internal processes and understand the context at each step, which differs significantly from traditional software debugging methods [31][32]. - The iterative process of developing agents is more complex, as developers cannot predict behavior before deployment, necessitating more rounds of refinement and adjustments [34][36]. Group 3: The Role of Data and Instructions - Existing software companies possess valuable data and APIs that can be leveraged in the agent era, but the ability to effectively utilize these assets will depend on new engineering approaches [37][38]. - The instructions on how to use data effectively are becoming increasingly important, as traditional methods of human execution are being automated through agents [38]. - The integration of domain-specific knowledge into agent systems is essential for their effectiveness, as seen in examples from the financial sector [38]. Group 4: Future of Agent Development - Memory capabilities in agents are anticipated to become a significant competitive advantage, allowing them to learn and improve over time [51][52]. - The development of user interfaces for long-horizon agents will likely require both synchronous and asynchronous management to handle tasks effectively [53][54]. - Code sandboxes are expected to become a critical component of agent capabilities, enabling safe execution and verification of scripts [56].

转自：InfoQ ，编译 | Tina 过去几十年，软件工程有一个稳定不变的前提：系统的行为写在代码里。工程师读代码，就能推断系 统在大多数场景下会怎么运行；测试、调试、上线，也都围绕"确定性"展开。但 Agent 的出现正在动 摇这个前提：在 Agent 应用里，决定行为的不再只是代码，还有模型本身——一个在代码之外运 行、带着非确定性的黑箱。你无法只靠读代码理解它，只能让它跑起来、看它在真实输入下做了什 么，才知道系统"到底在干什么"。 在播客中，LangChain 创始人 Harrison Chase 还把最近一波"能连续跑起来"的编程 Agent、Deep Research 等现象视为拐点，并判断这类"长任务 Agent"的落地会在 2025 年末到 2026 年进一步加 速。 这也把问题推到了台前：2026 被很多人视为"长任务 Agent 元年"，现有的软件公司还能不能熬过 去？就像当年从 on-prem 走向云，并不是所有软件公司都成功转型一样，工程范式一旦变化，就会 重新筛选参与者。长任务 Agent 更像"数字员工"——它不是多回合聊天那么简单，而是能在更长时间 里持续执行、反复试错、不断自 ...

Model Performance - Gemini 3 demonstrates extremely strong performance across various benchmarks, achieving state-of-the-art results in multiple areas [1] - Gemini 3 excels in tasks relevant to building agents, particularly in long horizon planning (Vending bench 2), terminal-based coding (Terminal Bench 2), and real-world contextual tasks like customer support (Sierra Tow Squared Bench) [2][3] Deep Agent Harness & Tool Utilization - The Deep Agent harness, an open-source tool, is used to test Gemini 3's agent-building capabilities, featuring built-in tools for planning, sub-agent delegation, and file system manipulation [3][4] - Gemini 3 effectively utilizes native tools within the Deep Agent harness, including file manipulation, planning, and sub-agent delegation, for tasks like research [18] - The agent successfully plans tasks, writes files, initiates sub-agents, analyzes results, updates to-dos, and generates final reports with citations [7][8] Research Task & Workflow - A research task is implemented using Gemini 3 within the Deep Agent harness, demonstrating the model's ability to perform complex tasks [5] - The research agent workflow involves creating to-dos, writing the research request to a file, initiating a sub-agent for research, analyzing results, and writing a final report [6][7] - The agent effectively uses a custom research sub-agent to isolate context, conduct in-depth research, and return results to the parent agent [15] Implementation & Customization - Gemini 3 can be easily integrated into existing workflows using Langchain and the Deep Agent harness [12][13] - The Deep Agent harness allows for customization through custom tools, instructions, and sub-agents, enabling users to tailor the agent to specific use cases [4][11] - The provided quick start repository offers instructions and code for running Gemini 3 with the Deep Agent harness, facilitating experimentation and customization [9][10]