2026-09-01
每日一文 · 长文精读
DoorDash’s Flux Runs 130,000 Engineering Tasks Through Cloud-Based Agents
DoorDash的Flux通过云端代理运行13万项工程任务
作者：Leela Kumili · InfoQ 原文
摘要：DoorDash将工程代理工作负载从开发者笔记本电脑迁移到Flux云平台，该平台在2026年单月自动化了13万项工程任务，每周支持超过2.5万次自动代码审查，使用超过300个剧本和每周超过1万次调用。Flux解决了本地执行的限制，基于四个平台原语：云沙箱、MCP网关、可重用剧本和调用面。沙箱使用Firecracker微虚拟机隔离，代理通过Agent网关访问内部系统。Flux最初用于自动代码审查，后扩展到CI分类、值班任务、维护工作流和工单驱动开发。
DoorDash has moved engineering agent workloads from developers’ laptops to its Flux cloud platform, which automated 130,000 engineering tasks in a single month in 2026.
DoorDash已将工程代理工作负载从开发者的笔记本电脑迁移到其Flux云平台，该平台在2026年单月内自动化了13万项工程任务。
The platform supports more than 25,000 automated code reviews each week, with more than 300 playbooks and more than 10,000 weekly invocations, allowing workflows to run unattended and in parallel.
该平台每周支持超过2.5万次自动代码审查，拥有超过300个剧本和每周超过1万次调用，使工作流能够无人值守且并行运行。
DoorDash developed Flux after encountering limitations with agent workloads running on individual laptops.
DoorDash在遇到代理工作负载在个人笔记本电脑上运行的局限性后，开发了Flux。
Local execution limits available CPU and memory, depends on the developer’s device remaining connected, and can give autonomous agents access to credentials and internal systems already available to the developer.
本地执行限制了可用的CPU和内存，依赖于开发者的设备保持连接，并且可能让自主代理访问开发者已有的凭据和内部系统。
DoorDash also said local execution makes it harder to monitor where agents are running, which systems they access, and on whose behalf they operate.
DoorDash还表示，本地执行使得更难监控代理在哪里运行、访问了哪些系统以及代表谁操作。
Radoslav Krehlik, security architect, wrote in a LinkedIn post about Flux,
安全架构师Radoslav Krehlik在关于Flux的LinkedIn帖子中写道：
Built after laptop-based agents ran into limits on power, uptime, and security.
在基于笔记本电脑的代理遇到计算能力、运行时间和安全性的限制之后构建。
Engineers can trigger the agents from Slack, GitHub, or scheduled jobs while DoorDash keeps full enterprise security guardrails.
工程师可以从Slack、GitHub或定时任务触发代理，同时DoorDash保持完整的企业安全护栏。
Flux is built around four platform primitives: cloud sandboxes, an MCP gateway, reusable playbooks, and invocation surfaces.
Flux围绕四个平台原语构建：云沙箱、MCP网关、可重用剧本和调用面。
Playbooks define the work, sandboxes provide isolated execution environments, the gateway controls access to internal systems, and invocation surfaces allow workflows to be started through Slack, GitHub, cron, the command line, or conversational interfaces.
剧本定义工作，沙箱提供隔离的执行环境，网关控制对内部系统的访问，调用面允许通过Slack、GitHub、cron、命令行或对话界面启动工作流。
The sandboxes use Firecracker micro virtual machines to isolate agent workloads.
沙箱使用Firecracker微虚拟机来隔离代理工作负载。
Each environment is provisioned with the repositories, development tools, secrets, and runtime dependencies required for a task.
每个环境都配备了任务所需的仓库、开发工具、密钥和运行时依赖项。
DoorDash reports a 95th percentile service level objective of less than five seconds for end-to-end sandbox setup, including starting the microVM, cloning repositories, installing build tools, and configuring the coding agent harness.
DoorDash报告端到端沙箱设置的95百分位服务等级目标低于五秒，包括启动微虚拟机、克隆仓库、安装构建工具和配置编码代理框架。
Agents access DoorDash’s internal systems through Agent Gateway, an in-house MCP gateway that provides scoped permissions and logs agent activity for auditing and policy enforcement.
代理通过Agent网关（一个内部MCP网关）访问DoorDash的内部系统，该网关提供范围限定的权限并记录代理活动以供审计和策略执行。
Playbooks, defined in YAML, specify the task, required tools, permissions, validation, and safety boundaries.
用YAML定义的剧本指定了任务、所需工具、权限、验证和安全边界。
They can combine agent-driven steps with deterministic code where predictable execution or validation is required.
它们可以将代理驱动的步骤与确定性代码相结合，适用于需要可预测执行或验证的场景。
Duy Nguyễn, Co-founder of TOP GROUP Vietnam, wrote in a LinkedIn post about Flux,
TOP GROUP Vietnam联合创始人Duy Nguyễn在关于Flux的LinkedIn帖子中写道：
Once agents handle background work, code review, internal tools, and concurrent execution, the hard problem shifts from model choice to control: identity, permissions, sandboxing, audit logs, quotas, and a real kill switch.
一旦代理处理了后台工作、代码审查、内部工具和并发执行，难题就从模型选择转向控制：身份、权限、沙箱、审计日志、配额和真正的终止开关。
The approach is part of a broader move toward cloud-based execution for coding agents.
这种方法是将编码代理转向云端执行的更广泛趋势的一部分。
GitHub supports both local and cloud sandbox environments for Copilot, with controls over filesystem, network, and system access.
GitHub为Copilot支持本地和云沙箱环境，并对文件系统、网络和系统访问进行控制。
Its cloud sandboxes provide isolated environments for agents to execute coding tasks without consuming local developer resources.
其云沙箱为代理提供隔离环境以执行编码任务，而不消耗本地开发者资源。
GitHub’s documentation provides additional details on its cloud-based execution model.
GitHub的文档提供了其基于云的执行模型的更多细节。
DoorDash initially used Flux for automated code review, which the company described as a frequent and measurable workflow.
DoorDash最初将Flux用于自动代码审查，该公司将其描述为一个频繁且可衡量的工作流。
It subsequently expanded the platform to CI triage, on-call tasks, maintenance workflows, and ticket-driven development.
随后，它将平台扩展到CI分类、值班任务、维护工作流和工单驱动开发。
DoorDash also changed its Slack integration from private channels to public threads so engineers could observe agent executions, review results, and see how other teams delegated work.
DoorDash还将其Slack集成从私密频道改为公开线程，以便工程师可以观察代理执行、审查结果，并了解其他团队如何委派工作。
阅读理解
1. What was the primary motivation for DoorDash to develop Flux?
2. Which component of Flux provides isolated execution environments for agent workloads?
3. According to the article, what change did DoorDash make to its Slack integration to improve transparency?