在 Codex 中运行实验的 GPT‑6 Astra 的助力下，Asana 在 GPT‑6.1 Sol 上优化了其浏览器代理的工作流，使其运行成本降低 76 倍，速度提升 5 倍。
Asana 通过 StackAI（在新窗口中打开） 帮助客户实现跨业务应用程序的工作自动化，该平台是其 收购（在新窗口中打开） 而来的。借助 StackAI，客户无需编写代码即可构建浏览网站、填写表单和收集信息的工作流。在 Asana 的规模下，这些工作流中的微小低效会不断累积。
Asana 的 StackAI 首席技术官 Frank Hidalgo 博士着手让浏览器代理运行得更快、更便宜。他指挥 Codex 中的 GPT‑6 Astra 调查该代理、测试改进并比较结果。他估计，这项工作若手工完成需要一到两个月，而实际只用了大约一周。
Asana 的 144 次运行研究（在新窗口中打开） 测试了 GPT‑6.1 Sol 和其他三个前沿模型，此处称为模型 A、B 和 C。在 GPT‑6.1 Sol 上得出的优化工作流平均估计模型成本为 0.47 美元，每次运行约四分钟，比最初在模型 B 上的生产设置便宜 76 倍、快 5 倍。
“这就是人类与代理团队在实践中的样子。一名工程师设定方向，GPT-6 Astra 运行实验，结果通过 Command 进入生产。这展示了 Asana 如何将人类与代理团队变为现实。”
——Arnab Bose，Asana 首席产品官
使用 GPT‑6 Astra 识别浏览器代理的低效之处
为了快速推进，Hidalgo 首先使用 Codex 中的 GPT‑6 Astra 来梳理代码库，并解释该代理如何构建每个模型请求。GPT‑6 Astra 发现，该代理缓存了固定指令和工具定义，但没有缓存其收集到的不断增长的页面文本和截图历史，因此每个请求都会以全价重新发送该历史。
该代理还几乎在每一步都丢弃较旧的截图并裁剪文本。每次编辑都会改变历史，因此仅缓存历史也无济于事，而丢失这些事实可能要求代理重新访问已经读过的页面。
从预计两个月的研究缩短到使用 GPT‑6 Astra 的一周
Hidalgo 审查了 GPT‑6 Astra 提出的修复方案，并选择其中三项进行测试：
将缓存扩展到代理的浏览历史
增加其可保留的文本量
按批次移除截图，而不是在每一步都移除
GPT‑6 Astra 先进行快速测试，以确定哪些变量重要。由于代码并非为受控实验而设计，它随后重构了代码，使一个前端和一个后端能够并行支持许多工作流，每个工作流都有自己的设置。
Astra 进行了完整研究：120,000 和 480,000 字符的历史预算，以及六种缓存和截图策略，每种都在四个模型上各测试三次（见下表）。表现最佳的策略允许截图累积到 20 张，然后再削减到最近一张。这使较早的历史在两次移除之间保持更长时间不变。结合更大的历史预算，它成为优化后的工作流。每种配置都执行相同任务：从公共演示目录中为 32 本书各收集六个字段，这代表了一些 Asana 客户在 StackAI 中运行的任务。
| 模型 | 描述 | 价格 |
|---|---|---|
| 模型 A | 来自另一家前沿实验室的较小、较便宜模型，于 2025 年秋季发布 | GPT‑6.1 Sol 价格的一半 |
| 模型 B | 最初用于生产的模型，来自与模型 A 相同的实验室，于 2026 年夏季发布 | 与 GPT‑6.1 Sol 价格相同 |
| 模型 C | 模型 B 的更新版本，于 2026 年秋季发布 | 与 GPT‑6.1 Sol 价格相同 |
| GPT‑6.1 Sol | OpenAI 的模型 |
GPT‑6 Astra 运行了这些工作流，并检查了请求、使用记录和输出，另有独立的模型会话审查了这项工作。每个会话的请求、数据轨迹和结果都记录在 Command（在新窗口中打开） 中，这是 Asana 的软件交付平台，因此团队之后可以审查完整研究。从 Command 中，研究结果被转化为工单，然后转化为拉取请求，变更随后进入生产。
“这手工完成需要我一到两个月。使用 Codex 中的 GPT-6 Astra，大约用了一周：我会在睡前设定一个 /goal，然后在早上审查结果。”
——Frank Hidalgo 博士，Asana 的 StackAI 首席技术官
将模型成本降至每次运行 0.50 美元以下
对于模型 B，优化将估计模型成本从至少 36.21 美元（一些原始运行在完成前达到了步数上限）降至每次运行 1.24 美元，减少了 29 倍。GPT‑6.1 Sol 上的优化工作流还要便宜 2.6 倍，为 0.47 美元。优化工作流中的每次运行都完成了任务并返回了正确答案。
3 次运行的均值。≥：基线包含达到上限的运行，因此其均值是下界。
右侧两个折叠与优化后的模型 B 进行比较。模型 B 在研究第 1 阶段运行，模型 C 和 Sol 6.1 在同一研究的第 2 阶段运行（虚线）。
仅在 GPT‑6.1 Sol 上，在更大的历史预算下，新的缓存和截图策略将成本降低了 4 倍，从每次运行 1.97 美元降至 0.47 美元。每次调用约便宜 3 倍，因为 89% 的输入来自缓存，价格为未缓存价格的 5%。运行也变得更快：在模型 B 上的原始设置中至少为 22.5 分钟，而在 GPT‑6.1 Sol 上的优化工作流中约为四分钟。
3 次运行的均值，SD 须线。≥：均值包含一次达到上限或未完成的运行，因此真实值至少这么大。
柱状图使用蓝色主题。请对照 480k 更大预算柱来解读缓存效果。
运行标记和 SD 须线是根据源图像进行的近似重建；底层运行值和标准差不可用。
3 次运行的均值，SD 须线。≥：均值包含一次达到上限或未完成的运行，因此真实值至少这么大。
柱状图使用蓝色主题。请对照 480k 更大预算柱来解读缓存效果。
运行标记和 SD 须线是根据源图像进行的近似重建；底层运行值和标准差不可用。
该调查还表明，历史管理会影响代理是否产生答案。给 GPT‑6.1 Sol 更多空间来保留其浏览历史，使产生答案的运行数量从较小历史预算下的 18 次中的 3 次增加到较大预算下的全部 18 次，且每次都给出正确答案。对 Hidalgo 而言，商业价值在于让客户能够使用更快、更强大的模型，同时保持运营成本可持续。
“过去，成本限制了我们能为客户提供哪些模型来处理这些工作负载。通过让智能体更高效，我们可以在降低运营成本的同时，为客户提供更好、更快的模型。”
——Frank Hidalgo，博士，Asana 的 StackAI 首席技术官
扩展实验和产品测试
Asana 已发布 StackAI 中浏览器导航的更改，并正在开发工具，使类似的实验更容易重复。随着时间的推移，团队计划将此测试纳入平台的评估中，以便客户和内部团队在配置其智能体时比较成本、运行时间和答案质量。
“发布速度不再是瓶颈；人类的注意力才是。我们正接近一个世界，在那里每位工程师都是领导一群智能体的产品经理。”
——Frank Hidalgo，博士，Asana 的 StackAI 首席技术官
Asana 现在正在 Codex 中使用 GPT‑6 Astra 在发布前测试产品功能：Astra 导航平台，尝试不同的输入，并向人工 QA 审查员报告错误。Hidalgo 将此视为新软件开发生命周期的基础，其中许多云智能体会话并行测试功能。
With GPT‑6 Astra in Codex running experiments, Asana optimized its browser agent’s workflow on GPT‑6.1 Sol to run 76x cheaper and 5x faster.
Asana helps customers automate work across business applications through StackAI(opens in a new window), a platform it acquired(opens in a new window). Using StackAI, customers can build workflows that navigate websites, fill out forms, and gather information without writing code. At Asana’s scale, small inefficiencies in these workflows add up.
Asana’s StackAI CTO, Frank Hidalgo, PhD, set out to make the browser agent faster and cheaper to run. He directed GPT‑6 Astra in Codex to investigate the agent, test improvements, and compare the results. Work he estimates would have taken one to two months by hand took about a week.
Asana’s 144-run study(opens in a new window) tested GPT‑6.1 Sol and three other frontier models, called here Models A, B and C. The optimized workflow that emerged on GPT‑6.1 Sol averaged $0.47 in estimated model costs and about four minutes per run, 76x cheaper and 5x faster than the original production setup on Model B.
“This is what teams of humans and agents look like in practice. An engineer set the direction, GPT-6 Astra ran the experiments, and the results went through Command to production. This demonstrates how Asana brings human and agent teams to life.”
—Arnab Bose, CPO at Asana
Identifying browser-agent inefficiencies with GPT‑6 Astra
To move quickly, Hidalgo started by using GPT‑6 Astra in Codex to map the codebase and explain how the agent built each model request. GPT‑6 Astra discovered that the agent cached its fixed instructions and tool definitions, but not the growing history of page text and screenshots it gathered, so every request resent that history at full price.
The agent also dropped older screenshots and trimmed text at nearly every step. Each edit altered the history, so caching the history alone would not have helped, and losing those facts could require the agent to revisit pages it had already read.
From an estimated two months of research to one week with GPT‑6 Astra
Hidalgo reviewed GPT‑6 Astra’s proposed fixes and selected three to test:
Extending caching to the agent’s browsing history
Increasing the amount of text it could retain
Removing screenshots in batches rather than at every step
GPT‑6 Astra began with quick tests to establish which variables mattered. Because the code was not designed for controlled experiments, it then refactored the code so one frontend and backend could support many workflows in parallel, each with its own settings.
Astra conducted the full study: history budgets of 120,000 and 480,000 characters and six caching and screenshot policies, each tested three times on each of the four models (see the table below). The best-performing policy allowed screenshots to accumulate to 20 before cutting back to the most recent one. This kept earlier history unchanged for longer stretches between removals. Combined with the larger history budget, it became the optimized workflow. Each configuration performed the same task: collecting six fields for each of 32 books from a public demo catalog, representative of what some Asana customers run in StackAI.
| Model | Description | Price |
|---|---|---|
| Model A | A smaller, less expensive model from another frontier lab, released Fall 2025 | Half the price of GPT‑6.1 Sol |
| Model B | The model originally used in production, from the same lab as Model A, released Summer 2026 | Same price as GPT‑6.1 Sol |
| Model C | An updated version of Model B, released Fall 2026 | Same price as GPT‑6.1 Sol |
| GPT‑6.1 Sol | OpenAI’s model |
GPT‑6 Astra ran the workflows and examined the requests, usage records and outputs, and separate model sessions reviewed the work. Every session’s requests, data traces and results were recorded in Command(opens in a new window), Asana’s software delivery platform, so the team could review the complete study afterward. From Command, the findings were turned into tickets, then pull requests, and the changes went to production.
“This would have taken me one to two months by hand. With GPT-6 Astra in Codex, it took about a week: I’d set a /goal before going to bed and review the results in the morning.”
—Frank Hidalgo, PhD, StackAI CTO at Asana
Bringing model cost below $0.50 per run
For Model B, the optimization reduced estimated model cost from at least $36.21 (some original runs reached the step limit before they finished) to $1.24 per run, a reduction of 29x. The optimized workflow on GPT‑6.1 Sol was 2.6x cheaper still, at $0.47. Every run in the optimized workflow completed the task and returned the correct answer.
Means of 3 runs. ≥: the baseline includes capped runs, so its mean is a lower bound.
The two right folds compare against Model B optimized. Model B ran in phase 1, Model C and Sol 6.1 in phase 2 of the same study (dotted line).
On GPT‑6.1 Sol alone, with the larger history budget, the new caching and screenshot policy reduced cost 4x, from $1.97 to $0.47 per run. Each call was about 3x cheaper, because 89% of the input came from cache at 5% of the uncached price. Runs also became faster: at least 22.5 minutes on the original setup on Model B, roughly four minutes with the optimized workflow on GPT‑6.1 Sol.
Mean of 3 runs, SD whiskers. ≥: mean includes a capped or unfinished run, so the true value is at least this large.
Bars use the blue theme. Read caching effects against the 480k larger-budget bar.
Run markers and SD whiskers are approximate reconstructions from the source image; underlying run values and standard deviations were not available.
Mean of 3 runs, SD whiskers. ≥: mean includes a capped or unfinished run, so the true value is at least this large.
Bars use the blue theme. Read caching effects against the 480k larger-budget bar.
Run markers and SD whiskers are approximate reconstructions from the source image; underlying run values and standard deviations were not available.
The investigation also showed how history management affected whether the agent produced an answer at all. Giving GPT‑6.1 Sol more room to retain its browsing history increased the number of runs that produced an answer from three of 18 with the smaller history budget to all 18 with the larger budget, each with the correct answer. For Hidalgo, the business value is giving customers access to faster, more capable models while keeping operating costs sustainable.
“Cost used to limit which models we could offer customers for these workloads. By making the agent more efficient, we can give customers a better, faster model while lowering our operating costs.”
—Frank Hidalgo, PhD, StackAI CTO at Asana
Scaling experimentation and product testing
Asana has released the changes to browser navigation in StackAI and is developing tools to make similar experiments easier to repeat. Over time, the team plans to incorporate this testing into the platform’s evaluations, so customers and internal teams can compare cost, runtime, and answer quality when configuring their agents.
“Shipping speed is no longer the bottleneck; human attention is. We’re close to a world where every engineer is a PM leading a fleet of agents.”
—Frank Hidalgo, PhD, StackAI CTO at Asana
Asana is now using GPT‑6 Astra in Codex to test product features before release: Astra navigates the platform, tries different inputs and reports bugs for human QA reviewers. Hidalgo views this as the foundation for a new software development lifecycle, with many cloud agent sessions testing features in parallel.
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文