Mind · In / Out · In · 事件

发布 Agents API

Introducing the Agents API

OpenAI Agents API 公测 · 2026-09-11

OpenAI 把执行环境外包给九家云商，只留下调度 agent 的那一层，等于亲口说明护城河在哪。

Indigo 的结论

OpenAI 把执行环境彻底外包，只握住调度层；价值往 agent 的运行框架和调度迁移，沙箱变成大路货。这对沙箱创业公司是坏消息：它们卖的是可替换的商品，定价权在上游。

怎么读这篇　OpenAI 官方加各家伙伴博客的合作公告，双方都在为自己说话，伙伴急着证明「该选我来跑 agent」。接线方式、双密钥、开通方式这些机制多方说法一致，可信度较高；Blaxel 被 Baseten 收购、DigitalOcean 邀请制预览是单一来源的二手综合，用之前先核原链接。

需要记住的几件事

  1. 八家接线方式一致：沙箱里跑 codex exec-server，用受限密钥向外连回 OpenAI，应用密钥永远不进沙箱。
  2. 各家卡位不同：Modal 抢 GPU，Blaxel 抢唤醒速度，E2B 抢隔离，Cloudflare 抢现有应用栈。
  3. 沙箱层被做成大路货，卖裸算力的一层也面临同样的结构：上游握着调度，下游拼价格。
  4. Blaxel 在进名单当天宣布被 Baseten 收购，独立的沙箱创业公司开始被推理层的公司吸收。

拆解 · 5 步

  1. 01

    把 Codex 的运行框架做成 API

    长时间运行的 agent 需要一套强的运行框架，加上能连跑几天的基础设施；这套东西现在开放公测。 读这一段原文 →

  2. 02

    八段客户证言充当效果证据

    评测分 0.71 到 0.85、延迟降 4 倍、每个案例成本降 60%、失败响应少 86%，全是客户自报。 读这一段原文 →

  3. 03

    一次调用建好 agent，运行框架归 OpenAI

    示例代码只指定任务、模型、工具和环境；OpenAI 托管运行框架，计算环境由开发者自己挑。 读这一段原文 →

  4. 04

    执行环境外包给九家伙伴

    列出 Blaxel、Cloudflare、Oracle 等一等集成伙伴，自家也备一套托管沙箱兜底。 读这一段原文 →

  5. 05

    留在自己手里的，是运行框架的迭代

    上下文压缩、工具搜索、子 agent 都在托管一侧持续升级；运行框架开源，API 本身不加价。 读这一段原文 →

什么会让我改口

上下文压缩、子 agent、工具搜索这些调度能力也被标准化成可替换的接口，或者哪家沙箱伙伴反过来长出了自己的调度层。

怎么读这篇

OpenAI 官方加各家伙伴博客的合作公告，双方都在为自己说话，伙伴急着证明「该选我来跑 agent」。接线方式、双密钥、开通方式这些机制多方说法一致，可信度较高；Blaxel 被 Baseten 收购、DigitalOcean 邀请制预览是单一来源的二手综合，用之前先核原链接。

拆解 · 5 步
  1. 把 Codex 的运行框架做成 API
  2. 八段客户证言充当效果证据
  3. 一次调用建好 agent，运行框架归 OpenAI
  4. 执行环境外包给九家伙伴
  5. 留在自己手里的，是运行框架的迭代
01

把 Codex 的运行框架做成 API

长时间运行的 agent 需要一套强的运行框架，加上能连跑几天的基础设施；这套东西现在开放公测。

用 Codex harness 搭建和运行云端 agent，全程由 OpenAI 托管。

我们把 Codex 和 ChatGPT for Work 推到全球数百万人手里，也因此摸清了长时间运行的 agent 要跑得好需要什么。好用的 agent 需要一套强的 harness，能管理上下文、高效调用工具、协调 subagent。它们还需要一套基础设施，能连着跑上好几天都稳，并且有可以读写文件、执行代码、保存中间结果的环境。

今天我们发布 Agents API 公测版，把驱动 Codex 的那套 harness 和基础设施，通过一个简单灵活的 API 交到开发者手里。

02

八段客户证言充当效果证据

评测分 0.71 到 0.85、延迟降 4 倍、每个案例成本降 60%、失败响应少 86%，全是客户自报。

客户怎么评价 Agents API

"用上 Agents API 之后，我们的评测分从 0.71 提到 0.85。API 里的 subagent 支持很好，大幅提速了我们的工作流。以前那套方案里观察和编排 subagent 相当麻烦，新 API 直接带来 4 倍的延迟下降。我们为此优化了很久，而 subagent 流程开箱就是一次巨大的提升。"

Ciridae CTO，Jack Weissenberger

"改造真实世界的生意，意味着把 AI 部署进形形色色的工作流。Agents API 提供 harness；环境、上下文和 UX 仍归我们自己。靠我们的 AI 平台 Nexus，我们现在能在几小时内把 agent 搭起来，跨行业落地，从住宅服务到建筑设计。"

Long Lake CTO，Rasmus Wissmann

"Agents API 让我们换了个思路去架构复杂的多步工作流。以前我们写 prompt 链、自己管一整套工具调用，现在可以直接在代码里用 agent，就像 Codex 在自己笔记本上跑那样。它已经帮我们解决了好几个问题，否则这些问题得自己造一套 agent 基础设施才行。"

WithCoverage 工程总监，Cole Striler

"把案件审核流程迁到 Agents API 之后，我们看到每个案件的成本下降 60%，延迟更低，token 效率明显改善，性能还保持原样。"

SafetyKit 技术团队成员，Bhavyansh Sabharwal

"测试里最打动我们的，是 Agents API 处理突发负载有多自然。我们可以把活扇出到几百个 agent 上，异步跑完，之后再收结果，不用在波峰之间一直养着闲置的基础设施。"

Dwelly 联合创始人兼 CTO，Dmitry Khanukov

"在金融服务业，赢得客户信任是关键。OpenAI 的 Agents API 让我们能造出更可靠的 agent，客户才敢在生产环境里用。把 agent harness 和沙箱分开之后，我们的 agent 失败响应减少了 86%。"

Hypha 主任工程师，Serhii Shchoholiev

"Agents API 在一个真实、活跃的代码仓库里完成了实现、独立评审、修复和真实浏览器验证。总体上，这个 agent 的工程质量非常强。"

deepsense.ai 资深 ML 工程师，Maks Operlejn

"在 Nash，我们部署了数千个长时间运行的 AI agent，管理着全球物流网络里数亿次配送。OpenAI 的 Agents API 给了我们需要的持久会话与编排层，让 agent 能在生产环境里连续运转，管好上下文、恢复和多步执行；Nash 则提供把它们接到物理世界的工具和执行环境。这让我们的 agent 能在跨越数小时甚至数天的复杂工作流里推理、行动、恢复和协作。这些 agent 就是生产基础设施，为我们的合作伙伴跑着关键的物流业务。"

Nash.ai 联合创始人兼 CTO，Aziz Alghunaim

03

一次调用建好 agent，运行框架归 OpenAI

示例代码只指定任务、模型、工具和环境；OpenAI 托管运行框架，计算环境由开发者自己挑。

一次 API 调用就能搭出云端 agent

用 Agents API，只要一次 API 调用，指定任务、模型、工具和环境，就能创建一个可上生产的 agent：

1import OpenAI from "openai";

3const client = new OpenAI();

5const session = await client.beta.agents.sessions.create({

11 server_label: "observability",

14 server_url: "https://observability.example.com/mcp",

18 multi_agent: { enabled: true, max_concurrent_subagents: 3 },

20 vault_ids: ["vault_YOUR_VAULT_ID"],

22 type: "openai_hosted",

23 capability_directories: ["/workspace/capabilities/skills"],

25 input:

26 "Investigate service-api's elevated 5xx rate over the last 30 minutes. " +

27 "Delegate deployment, error, and dependency analysis to subagents. " +

28 "Save findings, evidence, and recommended mitigation in /workspace/outputs.",

harness 由 OpenAI 托管和维护。agent 的计算环境由开发者自己挑：可以放在 OpenAI 托管的沙箱里，可以放在自家基础设施上，也可以交给我们的沙箱伙伴之一。Agents API 把我们优化过的 agent harness 和基础设施做成地基，让开发者能专心打磨那些让自家 agent 与众不同的工具、知识和工作流。

Agents API 用驱动 Codex 的同一套 harness 和基础设施，来驱动开发者自己的 agent。

04

执行环境外包给九家伙伴

列出 Blaxel、Cloudflare、Oracle 等一等集成伙伴，自家也备一套托管沙箱兜底。

挑一个合适的 agent 环境

不同负载需要不同的计算、存储和部署方式。Agents API 让开发者挑一个适合自家应用的沙箱。

我们正在与生态伙伴合作，包括 Blaxel、Cloudflare、Daytona、DigitalOcean、E2B、Modal、Oracle、Runloop 和 Vercel，为各种需求提供一等集成：

完全托管的环境，或者部署在自家 VPC 里

特定的文件与密钥存储机制

不同的 CPU、GPU 和内存配置，性能、冷启动和成本曲线各不相同，可以按公司的工作流来配。

Agents API 与主流生态伙伴提供一等集成。

OpenAI 托管沙箱

给那些想快速上手、高效扩容的开发者，我们同时推出 OpenAI 托管沙箱。它用的就是驱动 Codex 和 ChatGPT 的那套沙箱基础设施。

沙箱由 OpenAI 开通和管理，给 agent 一个安全、够快的环境去执行代码、处理文件、产出成果。这些沙箱可以灵活装上你的文件、软件包、skill 和插件，让 agent 拿到完成任务所需的一切。

05

留在自己手里的，是运行框架的迭代

上下文压缩、工具搜索、子 agent 都在托管一侧持续升级；运行框架开源，API 本身不加价。

跟着 Codex harness 一起进化

想吃上模型的新能力，往往意味着要重做 harness，把本该用来打磨应用的时间耗掉。Agents API 在每次模型发布时提供带版本的能力接入。我们和模型一起维护并持续改进 harness，让 agent 每次升级都能拿到更好的表现。比如最近对 harness 的改进包括：

让 agent 在长会话里持续工作

为了撑住模型连续工作数小时，我们做了上下文管理，帮 agent 把相关信息带过更长的会话。当会话逼近上下文上限时，Agents API 会自动压缩（compact）早前的上下文，保住 agent 继续干活所需的信息。开发者不用自己实现压缩逻辑，就能搭出跨越多个上下文窗口的工作流。

帮 agent 高效地用上更多工具

Agents API 帮 agent 找到对的工具并高效使用。工具搜索（tool search）按需加载相关的工具定义，既省 token 和成本，又不破坏模型的缓存。工具就位之后，程序化工具调用（programmatic tool calling）让 agent 并行发起调用、串起相关操作、在代码里筛选或合并结果，从而处理大批量数据，只把相关结果带回上下文。Agents API 支持 MCP、自定义函数，以及网页搜索这类内置工具。

5 "server_label": "openai_docs",

8 "server_url": "https://developers.openai.com/mcp"

让 agent 用 subagent 把活并行拆开

有了多 agent 支持，Agents API 能把复杂任务拆成互不依赖的小块，派给并行工作的 subagent。每个 subagent 维护自己的上下文，因此能专注在分到的那份活上，主 agent 则负责协调并把结果汇总起来。这能给受益于并行的研究、分析和编码任务提速，而且不用自己造一套编排。

2 "model": "gpt-6-astra",

5 "max_concurrent_subagents": 3,

一个开源的地基

Agents API 由开源的 Codex harness 驱动，开发者能看清协调模型调用、工具和上下文的那套核心逻辑。用 Agents API 时，OpenAI 负责运行和维护这套 harness，开发者则可以查阅并研究它的公开代码库。

开始搭建

Agents API 今天起对所有开发者开放公测。用 Agents API 不额外收费——只按 agent 消耗的 token 和工具付费，具体见我们的定价页。

想了解更多可以看 Agents API 总览，或者跟着快速上手做一遍，把 Codex 背后的这套 harness 带进自己的 agent。

公测期间我们会根据反馈快速迭代，朝正式可用推进。告诉我们哪里好用、哪里别扭、把 agent 跑到生产还缺什么。

判断收口延伸

Indigo 的结论

OpenAI 把执行环境彻底外包，只握住调度层；价值往 agent 的运行框架和调度迁移，沙箱变成大路货。这对沙箱创业公司是坏消息：它们卖的是可替换的商品，定价权在上游。

需要记住的几件事

  1. 八家接线方式一致：沙箱里跑 codex exec-server，用受限密钥向外连回 OpenAI，应用密钥永远不进沙箱。
  2. 各家卡位不同：Modal 抢 GPU，Blaxel 抢唤醒速度，E2B 抢隔离，Cloudflare 抢现有应用栈。
  3. 沙箱层被做成大路货，卖裸算力的一层也面临同样的结构：上游握着调度，下游拼价格。
  4. Blaxel 在进名单当天宣布被 Baseten 收购，独立的沙箱创业公司开始被推理层的公司吸收。

放回主线

证实

你拥有的不是模型　OpenAI 把执行环境外包、只握调度层，是这条判断最直接的平台方自证：模型和沙箱都在变成大路货。

补充

验证不可压缩　沙箱负责运行和检验 agent 的输出；这批伙伴真正比的，是能不能安全、可复现地运行不可信的代码。

补充

Lin Qiao（Fireworks）：post-training 是你保住品味的方式　同一主题的两面：一边说护城河在把判断练进权重，这边说 OpenAI 把护城河锁在调度层。

什么会让我改口

上下文压缩、子 agent、工具搜索这些调度能力也被标准化成可替换的接口，或者哪家沙箱伙伴反过来长出了自己的调度层。

读完了。Indigo 对这篇的判断在这两处：

Mind · In / Out · In · Event

Introducing the Agents API

openai.com · 2026-09-11

OpenAI hands code execution to nine cloud partners and keeps only the layer that runs the agent: it is telling us where its moat is.

Indigo's conclusion

OpenAI outsources execution entirely and keeps orchestration. Value moves to the harness and orchestration, and sandboxes become commodities. Bad news for sandbox startups: they sell something swappable, and pricing power sits upstream.

How to read this　A partnership announcement from OpenAI plus each partner's own blog; everyone is talking their book, and the partners are eager to prove “pick us to run your agents”. The mechanics (wiring, the two keys, provisioning) agree across sources and are fairly reliable. Blaxel's acquisition by Baseten and DigitalOcean's invite-only preview come from a single second-hand summary; check the original links before acting.

What to remember

  1. All eight partners wire it the same way: codex exec-server runs in the sandbox, a restricted key connects back out to OpenAI, and the application key never enters the sandbox.
  2. Modal competes on GPUs, Blaxel and Runloop on millisecond wake-up, E2B and Daytona on isolation, Cloudflare and Vercel on existing app stacks.
  3. With sandboxes commoditized, the raw-compute layer faces the same structure: orchestration held upstream, price competition downstream.
  4. Blaxel announced its acquisition by Baseten the day it made the list: independent sandbox startups are being absorbed by inference players.

Breakdown · 5 steps

  1. 01

    Codex's harness, offered as an API

    Long-running agents need a strong harness plus infrastructure that can run for days. That package is now in public beta. Read this part →

  2. 02

    Eight customer testimonials stand in as evidence

    Eval scores of 0.71 to 0.85, latency down 4x, cost per case down 60%, 86% fewer failed responses: all self-reported by customers. Read this part →

  3. 03

    One call builds an agent; OpenAI keeps the harness

    The sample code specifies only the task, model, tools and environment. OpenAI hosts the harness; developers choose where the code runs. Read this part →

  4. 04

    Execution handed to nine partners

    Blaxel, Cloudflare, Oracle and others are listed as first-class integrations, with OpenAI's own hosted sandbox as a fallback. Read this part →

  5. 05

    What OpenAI keeps: improving the harness

    Context compaction, tool search and subagents keep improving on the hosted side. The harness is open source, and the API adds no markup. Read this part →

What would change my mind

orchestration features such as compaction, subagents and tool search also become standardized, swappable interfaces, or a sandbox partner grows its own orchestration layer.

How to read this

A partnership announcement from OpenAI plus each partner's own blog; everyone is talking their book, and the partners are eager to prove “pick us to run your agents”. The mechanics (wiring, the two keys, provisioning) agree across sources and are fairly reliable. Blaxel's acquisition by Baseten and DigitalOcean's invite-only preview come from a single second-hand summary; check the original links before acting.

Breakdown · 5 steps
  1. Codex's harness, offered as an API
  2. Eight customer testimonials stand in as evidence
  3. One call builds an agent; OpenAI keeps the harness
  4. Execution handed to nine partners
  5. What OpenAI keeps: improving the harness
01

Codex's harness, offered as an API

Long-running agents need a strong harness plus infrastructure that can run for days. That package is now in public beta.

Build and run cloud agents with the Codex harness, fully managed by OpenAI.

As we’ve scaled Codex and ChatGPT for Work to millions of people around the world, we’ve learned what it takes to make long-running agents work well in practice. Useful agents need a powerful harness that manages context, uses tools efficiently, and coordinates subagents. They also need infrastructure that keeps them running reliably for days, with environments where they can work with files, run code, and save intermediate results.

Today, we’re introducing the Agents API in public beta, bringing that same harness and infrastructure that powers Codex to developers through a simple, flexible API.

02

Eight customer testimonials stand in as evidence

Eval scores of 0.71 to 0.85, latency down 4x, cost per case down 60%, 86% fewer failed responses: all self-reported by customers.

What our customers are saying about Agents API

“With the Agents API, our evaluation score went from 0.71 to 0.85. The subagent support in the API is great and drastically sped up our workflow. Previously it was pretty cumbersome to observe and orchestrate subagents in our old setup but the new APIs gave us a 4x latency reduction. We spent a long time trying to optimize for this and the subagent flows were a huge out-of-the-box lift.”

Jack Weissenberger, CTO, Ciridae

“Transforming real-world businesses means deploying AI into workflows of every shape. Agents API supplies the harness; the environment, context, and UX stay ours. With our AI platform Nexus we now stand up agents in hours across industries, from residential services to architecture.”

Rasmus Wissmann, CTO, Long Lake

“The Agents API has enabled us to think differently about how we can architect complex, multi-step workflows. We used to write prompt chains and manage our own set of tool calls, but now we can use agents directly in our code much like how Codex works on your laptop. It’s already helped us solve several problems that would’ve otherwise required us to build custom agent infrastructure.”

Cole Striler, Director of Engineering, WithCoverage

“After migrating our case review workflow to the Agents API, we saw a 60% reduction in cost per case, lower latency, and significantly improved token efficiency while maintaining existing performance.”

Bhavyansh Sabharwal, Member of Technical Staff, SafetyKit

“What stood out in our testing was how naturally the Agents API handled bursty workloads. We could fan out work across hundreds of agents, run them asynchronously, and collect the results later, without keeping infrastructure idle between peaks.”

Dmitry Khanukov, Co-founder & CTO, Dwelly

“Earning customers’ trust is critical in financial services. OpenAI’s Agents API enables us to build more reliable agents, giving customers the confidence to use them in production. By separating the agent harness from the sandbox, we reduced failed agent responses by 86%.”

Serhii Shchoholiev, Lead Engineer, Hypha

“The Agents API handled the implementation, independent review, remediation, and real-browser validation in a real, active repository. Overall, the agent’s engineering quality was very strong.”

Maks Operlejn, Senior ML Engineer, deepsense.ai

“At Nash, we deploy thousands of long-running AI agents that manage hundreds of millions of deliveries across global logistics networks. OpenAI’s Agents API gives us the durable session and orchestration layer we need for agents operating continuously in production managing context, recovery, and multi-step execution, while Nash provides the tools and execution environment that connect them to the physical world. This lets our agents reason, act, recover, and collaborate across complex workflows that can span hours or days. These agents are production infrastructure running mission-critical logistics operations for our partners.”

Aziz Alghunaim, Co-founder & CTO, Nash.ai

03

One call builds an agent; OpenAI keeps the harness

The sample code specifies only the task, model, tools and environment. OpenAI hosts the harness; developers choose where the code runs.

Build cloud agents with a single API call

With the Agents API, you can create a production-ready agent in a single API call by specifying the task, model, tools, and environment:

1import OpenAI from "openai";

3const client = new OpenAI();

5const session = await client.beta.agents.sessions.create({

11 server_label: "observability",

14 server_url: "https://observability.example.com/mcp",

18 multi_agent: { enabled: true, max_concurrent_subagents: 3 },

20 vault_ids: ["vault_YOUR_VAULT_ID"],

22 type: "openai_hosted",

23 capability_directories: ["/workspace/capabilities/skills"],

25 input:

26 "Investigate service-api’s elevated 5xx rate over the last 30 minutes. " +

27 "Delegate deployment, error, and dependency analysis to subagents. " +

28 "Save findings, evidence, and recommended mitigation in /workspace/outputs.",

OpenAI hosts and maintains the harness. You choose the agent’s compute environment: in an OpenAI-managed sandbox, on your own infrastructure, or with one of our sandbox partners. The Agents API gives you a strong foundation for building agents on top of our optimized agent harness and infrastructure, so you can focus on the tools, knowledge, and workflows that make your agent unique.

Agents API powers your agents with the same harness and infrastructure behind Codex.

04

Execution handed to nine partners

Blaxel, Cloudflare, Oracle and others are listed as first-class integrations, with OpenAI's own hosted sandbox as a fallback.

Choose your agent environment

Different workloads need different compute, storage, and deployment options. The Agents API lets you choose a sandbox that fits your application.

We’re partnering with ecosystem providers, including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel, to provide first-class integrations for a range of needs:

Fully managed environments or deployments within your VPC

Specific file and secret storage mechanisms

Different CPU, GPU, and memory configurations, with performance, cold-start, and cost profiles to match your company’s workflow.

The Agents API offers first-class integrations with popular ecosystem providers.

OpenAI hosted sandboxes

For developers who want to get started quickly and scale efficiently, we’re also introducing the OpenAI hosted sandbox. This leverages the same sandboxing infrastructure that powers Codex and ChatGPT.

OpenAI provisions and manages the sandbox, giving your agent a secure and performant environment to run code, work with files, and produce artifacts. These sandboxes can be flexibly configured with your files, packages, skills and plugins to give the agent what it needs to complete the task.

05

What OpenAI keeps: improving the harness

Context compaction, tool search and subagents keep improving on the hosted side. The harness is open source, and the API adds no markup.

Build with an evolving Codex harness

Taking advantage of new model capabilities often means reworking your harness, taking valuable time away from improving your application. The Agents API provides versioned access to these capabilities with each model launch. We maintain and continuously improve the harness alongside our models, helping your agents get better performance from every upgrade. For example, recent improvements to the harness include:

Keep agents working across long sessions

To support models working for hours, we’ve built context management that helps agents carry relevant information across longer sessions. The Agents API automatically compacts earlier context as a session approaches its context limit, preserving information the agent needs to continue. Developers can build workflows that span multiple context windows without implementing their own compaction logic.

Help agents efficiently use more tools

The Agents API helps agents find the right tools and use them efficiently. Tool search loads relevant tool definitions as needed, helping reduce token usage and cost while preserving the model’s cache. Once tools are available, programmatic tool calling lets agents run calls in parallel, chain related operations, and filter or combine results in code so they can work through large volumes of data while bringing only the relevant results back into context. The Agents API supports MCP, custom functions, and built-in tools like web search.

5 "server_label": "openai_docs",

8 "server_url": "https://developers.openai.com/mcp"

Let agents parallelize work with subagents

With multi-agent support, the Agents API can break complex tasks into independent pieces and delegate them to subagents that work in parallel. Each subagent maintains its own context, helping it stay focused on its assignment, while the main agent coordinates their work and brings the results together. This can speed up research, analysis, and coding tasks that benefit from parallel work, without requiring you to build your own orchestration.

2 "model": "gpt-6-astra",

5 "max_concurrent_subagents": 3,

An open-source foundation

The Agents API is powered by the open-source Codex harness, giving developers visibility into the core logic that coordinates model calls, tools, and context. With the Agents API, OpenAI operates and maintains that harness while developers can inspect and learn from its public codebase.

Start building

Agents API is available in public beta today to all developers. There are no additional fees for using the Agents API – you simply pay for the tokens and tools your agents use, as outlined on our pricing page.

Explore the Agents API overview to learn more, or follow the quickstart to get started and bring the harness behind Codex into your own agents.

During the public beta, we’ll iterate quickly based on your feedback as we work toward general availability. Let us know what’s working, where you’re running into friction, and what you need to build and run your agents in production.

Where Indigo landsFurther

Indigo's conclusion

OpenAI outsources execution entirely and keeps orchestration. Value moves to the harness and orchestration, and sandboxes become commodities. Bad news for sandbox startups: they sell something swappable, and pricing power sits upstream.

What to remember

  1. All eight partners wire it the same way: codex exec-server runs in the sandbox, a restricted key connects back out to OpenAI, and the application key never enters the sandbox.
  2. Modal competes on GPUs, Blaxel and Runloop on millisecond wake-up, E2B and Daytona on isolation, Cloudflare and Vercel on existing app stacks.
  3. With sandboxes commoditized, the raw-compute layer faces the same structure: orchestration held upstream, price competition downstream.
  4. Blaxel announced its acquisition by Baseten the day it made the list: independent sandbox startups are being absorbed by inference players.

Back on the long-running theses

confirms

You don't own the model　OpenAI outsourcing execution and keeping orchestration is the most direct confirmation from a platform: models and sandboxes are both becoming commodities.

adds to

Verification can't be compressed　Sandboxes run and check agent output; what these partners really compete on is running untrusted code safely and reproducibly.

adds to

Lin Qiao (Fireworks): post-training is how you keep your taste　Two sides of one theme: there the moat is judgment trained into the weights; here OpenAI locks its moat in orchestration.

What would change my mind

orchestration features such as compaction, subagents and tool search also become standardized, swappable interfaces, or a sandbox partner grows its own orchestration layer.

Finished. Indigo's take on this piece is in two places: