Third Moment Research 如何运行 10，000 个代理 36+ 小时

每家公司都有自己的工作流程、工具、边缘情况和成功定义。Third Moment Research（在新窗口中打开）认为，最好的智能体应该擅长处理特定工作，并且随着使用次数增加而不断提升能力。Every company has its own workflows, tools, edge cases, and definitions of success. ⁠(opens in a new window)Third Moment Research⁠(opens in a new window) believes the best agents should be really good at that specific work and should improve their capabilities the more they’re used.

该公司正围绕这一理念打造一个能够自我改进的 AI 系统，利用 OpenAI 的最新模型来理解智能体的行为、识别反复出现的弱点，并围绕这些问题自动生成数千个合成场景。团队利用这些场景改进智能体，并随着新问题不断出现而重复这一流程。The company is building a self-improving AI system around that idea, using OpenAI’s latest models to understand agent behavior, identify recurring weaknesses, and automatically create thousands of synthetic scenarios around those problems. The team uses those scenarios to improve agents, repeating the process as new issues emerge.

Codex、GPT‑6 Astra、GPT‑6 Sol 和 GPT‑6.1 Sol 为系统的研究与计算工作提供支持，使 Third Moment 能够同时运行 10,000 个智能体，开展持续 36 至 48 小时的调查。该公司还使用 Daybreak 测试智能体工作环境的安全性。Codex, GPT‑6 Astra, GPT‑6 Sol, and GPT‑6.1 Sol support the system’s research and computational work, enabling Third Moment to run 10,000 agents at once on investigations lasting 36-48 hours. The company also uses Daybreak to test the security of the environments where agents work.

Third Moment 的早期应用之一是量化金融领域，因为这里有明确的成功衡量标准。这个领域为研究智能体如何工作以及它们在哪些方面遇到困难提供了一个具体场景。团队使用 Codex 来发现并解决研究环境中的问题。One of Third Moment’s early applications is in the field of quantitative finance given the clear measure of success. This domain provides a concrete setting to study how agents work and where they struggle. The team uses Codex to find and address problems in the research environment.

有一次，它使用当时最新的 GPT 模型 GPT‑5.6，通过 Codex 分析了数千条记录下来的仿真轨迹（智能体的行动及环境的响应），以及时间戳和代码库。In one instance, it used Codex with GPT‑5.6, the latest GPT model at the time, to analyze thousands of recorded simulation trajectories (the agents’ actions and the environment’s responses) alongside timestamps and the codebase.

Codex发现了一个瓶颈：在一年份的美国股票订单级数据上对交易策略进行回测需要数小时。等待期间，智能体反复轮询任务是否完成，浪费了令牌，还让上下文被状态检查信息填满。Codex uncovered a bottleneck: backtesting a trading strategy across a year of order-level US equity data took several hours. While waiting, agents repeatedly polled for completion, wasting tokens and filling their context with status checks.

对回测进行性能分析并分析其输入和用法后，Codex 找到了一项缓存优化和一种内存打包技术。它尝试了多种方案，并将修复部署到生产环境，将回测耗时从约 3 小时缩短至约 3 到 5 分钟。这减少了 token 使用量和上下文污染，为智能体留出了更多 token 预算用于后续实验。After profiling the backtest and analyzing its inputs and usage, Codex identified a caching optimization and a memory-packing technique. It experimented with solutions and deployed the fix to production, reducing the backtest from approximately 3 hours to approximately 3 to 5 minutes. This reduced token use and context pollution, freeing more of the agents’ token budget for further experiments.

Third Moment 为每个智能体提供完成特定工作所需的工具、信息和环境。在量化研究中，团队会在 Docker 化环境中运行默认的 Codex harness。智能体可以访问 1 TB 的数据、文档和研究代码库，其中包含用于特征工程、GPU 支持的模型训练、交易策略和模拟的模块。Third Moment gives each agent the tools, information, and environment it needs for a specific job. For quantitative research, the team runs the default Codex harness inside a Dockerized environment. Agents receive a terabyte of data, documentation, and a research repository with modules for feature engineering, GPU-supported model training, trading strategies, and simulation.

GPT‑6 Astra、GPT‑6 Sol 和 GPT‑6.1 Sol 负责推理，而 Codex 框架则为智能体提供工具，使其能够采取行动、制定计划、设定目标、生成子智能体，并在工作持续进行时压缩上下文。GPT‑6 Astra, GPT‑6 Sol, and GPT‑6.1 Sol provide the reasoning, while the Codex harness gives agents tools to act, create plans, set objectives, spawn subagents, and compact context as work continues.

为了解决代理在完成任务前停止的问题，Third Moment 将 Codex 视为由 GPT‑6 Sol 或 GPT‑6.1 Sol 驱动的外层执行框架中的子代理。当任务未完成时，外层执行框架会检查执行轨迹，并决定是重试、终止作业，还是升级至人工审核。To address agents that stop before completing a task, Third Moment treats Codex as a subagent within an outer harness powered by GPT‑6 Sol or GPT‑6.1 Sol. The outer harness examines the trajectory when a task is left unfinished and determines whether to retry, terminate the job, or escalate to human review.

在这一研究环境中，智能体致力于改进由 Third Moment 工程师构建的基准交易系统。他们在历史数据上测试交易策略，评估改进究竟有意义还是只是噪声，并根据每次实验的结果决定下一步实验。Adam 表示，他们的工作已经改进了基准交易系统。In that research environment, agents work to improve a baseline trading system built by Third Moment’s engineers. They test trading strategies on historical data, assess whether an improvement is meaningful or noise, and use each result to choose the next experiment. According to Adam, their work has improved the baseline trading system.

在 GPT‑5.5 之前，Adam 表示，模型只能执行用户指定的实验。到了 GPT‑5.5 和 GPT‑5.6，使用 Codex harness 的模型可以根据实验结果设计后续实验。如今，借助 GPT‑6 Astra、GPT‑6 Sol 和 GPT‑6.1 Sol，智能体可以通过与子智能体并行运行的数百项实验，调查问题的根本原因。Before GPT‑5.5, Adam says models could only execute experiments prescribed by a user. With GPT‑5.5 and GPT‑5.6, models working with the Codex harness could use results to design follow-up experiments. Now with GPT‑6 Astra, GPT‑6 Sol, and GPT‑6.1 Sol, agents can investigate root causes through hundreds of experiments running in parallel with subagents.

这种演进帮助智能体探索团队此前未曾想到的研究方向。在一次研究过程中，GPT‑6.1 Sol 发现，在某家交易所，非显示交易的影响不容忽视，因为这类交易占据的成交量份额大得出乎意料。Adam 表示：“这是一个不显而易见的发现，我们的研究团队最初并未考虑到这一点。”That progression has helped agents explore research directions the team hadn’t anticipated. In one session, GPT‑6.1 Sol identified the impact of non-displayed trades on an exchange where those trades accounted for an unexpectedly large share of volume. “This was an unobvious result that our research team did not initially consider,” says Adam.

Third Moment 从优化、机器学习系统和量化研究等方面评估模型。它会在最高推理设置下测试前沿版本，评估模型浏览代码仓库、编写高效代码和持续开展研究的能力。目前的测试也包括 GPT‑6 Luna。Third Moment evaluates models across optimization, machine learning systems, and quantitative research. It tests frontier releases at their highest reasoning settings, assessing their ability to navigate repositories, write efficient code, and sustain an investigation. Current testing also includes GPT‑6 Luna.

一项基准测试要求智能体在保持输出不变的同时，降低实时记分牌服务的更新延迟。团队通过将专家编写的解决方案的平均运行时间除以模型解决方案的平均运行时间来衡量加速比，前提是结果正确。据 Adam 称，GPT‑6 Astra 实现了最大的加速比。One benchmark asks agents to reduce a live scoreboard service’s update latency while preserving its outputs. The team measures speedup by dividing the average runtime of an expert-crafted solution by that of the model’s solution on the same hardware, subject to correctness. GPT‑6 Astra delivered the biggest speedup, according to Adam.

对于有针对性的工作，包括检查模拟系统中是否存在特定安全漏洞，团队会使用 GPT‑5.6 Terra 这类效率更高的模型。Adam 认为，在深度研究中，更强的推理能力值得更高的成本：“在量化金融领域，发现某些有意思信号的价值极高。”For targeted work, including checks for specific security weaknesses in the simulation system, the team uses more efficient models like GPT‑5.6 Terra. For deep research, Adam believes stronger reasoning can justify higher cost: “In the space of quantitative finance, the value of finding some interesting signal is extremely high.”

随着 10,000 个智能体自主运行，Third Moment 需要一种方法，大规模测试其研究环境的安全性。团队使用 Daybreak 网络安全平台，在模拟启动前自动执行这些测试。With 10,000 agents running autonomously, Third Moment needs a way to test the security of its research environments at scale. The team uses the Daybreak cybersecurity platform to automate that testing before simulations launch.

这些测试会检查智能体是否能够突破其研究环境并访问更广泛的网络。Adam 说：“借助 Daybreak，我们可以让智能体自行对模拟环境进行渗透测试，并测试其基础设施和安全性。”The tests check whether an agent could break out of its research environment and access the wider network. “By using Daybreak, we’re able to have agents themselves do penetration testing on the simulation and test the infrastructure and security of it,” says Adam.

Third Moment 的愿景还涵盖企业每天开展的专业工作。对零售商而言，这可能意味着在库存短缺或供应商延迟时，做出更好的供应链决策。对工程团队而言，这可能意味着在大型代码库中找出性能瓶颈。对依赖大量文档开展业务的企业而言，这可能意味着发现一份文件中的承诺与另一份文件记录的结果不一致之处。Third Moment’s ambition extends to the specialized work companies do every day. For a retailer, that could mean making better supply chain decisions when inventory is short or a supplier is delayed. For an engineering team, it could mean finding performance bottlenecks in a large codebase. For a document-heavy business, it could mean identifying where commitments in one document don’t match outcomes recorded in another.

基于 GPT‑6 Astra、GPT‑6 Sol 和 GPT‑6.1 Sol 的推理能力，Third Moment 计划将规模从单项调查扩展到数十万个协同工作的子代理。其路线图设想投入数万亿个 token，支持由数十个相互依赖的实验组成的链条，旨在利用这些研究，在新挑战出现时自主改进代理。Building on the reasoning capabilities of GPT‑6 Astra, GPT‑6 Sol, and GPT‑6.1 Sol, Third Moment plans to expand from individual investigations to hundreds of thousands of collaborating subagents. Its roadmap calls for trillions of tokens supporting chains of dozens of dependent experiments, with the goal of using that research to autonomously improve agents as new challenges arise.

对 Adam 来说，这能让人们更有能力去塑造智能体的学习和工作方式：“人类可以处于更具杠杆效应的位置，思考如何开展研究、如何优化研究本身，而不是亲自做研究。”For Adam, this would give people more capacity to shape how agents learn and work: “Humans can be in a much more leveraged position, where they can think about how to do research and optimize the research itself instead of actually doing the research.”

继续阅读Keep reading