Xirp：Spotify 打造的智能体开发环境

我接触过的每一位 CTO，都已经不再讨论是否采用的问题。他们的工程组织每天都在使用 AI 编程智能体。但其所处的环境持续变化：新模型、新智能体、不断上升的成本、每季度都可能调整的定价结构，以及一夜之间就能取得进展的能力。

真正的挑战已不再只是提升速度，而是确保团队能够在任意时刻灵活采用价值最大的工具，同时不积累供应商依赖，也不会在规模扩大时让成本不断叠加。

在 Spotify，数千名工程师分布在数百个团队中工作，这种张力决定了我们构建开发基础设施的方式。我们的工程师需要自主选择最适合其工作的工具，而“最适合”的标准正在迅速变化。但没有结构的自主性会造成碎片化、重复投入和不一致的结果。如何调和这种张力，正是推动我们采取这一方法的原因。

现在回看，这一演进过程是可以预见的。团队最初使用单个智能体：一个会话、一名开发者、一个任务。随后，采用速度开始加快。

工程师，以及后来的非工程人员，开始在不同会话、工具、代码仓库和分支中并行编排多个智能体。上下文在一个会话中积累，却需要在另一个会话中使用。组织知识被分散到个人配置中——CLAUDE.md 文件、定制化 MCP 设置、个人提示词库——使得这些上下文更难跨团队和工作流流转。

由此产生的混乱影响了生产力。PR 数量增加了，但返工、不一致和 Token 浪费也随之增加。一个会话中的智能体会耗费计算周期，重新发现另一个会话其实已经解决的问题。

此外，工程师花在重建上下文上的精力，几乎与构建新能力的精力一样多。在组织层面，如何在数千个并发会话中建立更强的可见性，也变得越来越重要。

这对我们而言并不是一类新问题。十年前，数千名工程师在数千项服务上工作的复杂性，就产生了同类别的碎片化。因此，我们构建了 Backstage：一个统一、结构化的事实来源，用于描述系统如何相互连接，包括依赖关系图、组件元数据、所有权和架构决策。Backstage 成为了我们内部开发者平台的基础；随后我们又构建了 Portal，使任何组织都能获得同样的收益，而无需自行管理基础设施。

AI 编程智能体为这种组织复杂性引入了一个全新的维度。它需要专门的解决方案。

Xirp 是一个由具体工程需求催生、供应商中立的智能体开发环境。

Xirp 的构建目标，是帮助我们的开发者跨多个运行框架（Claude Code、Gemini CLI、Codex 等）管理数十个并发智能体会话，使协调 50 个以上并行会话成为可行之事。每个会话都在独立的 worktree 中运行，使数十个智能体能够在同一代码库上并发工作而互不干扰。上下文与任何单一智能体或运行框架解耦；即使在项目进行中切换工具，完整的工作状态也会随之保留。供应商中立并非附加特性，而是直接源自我们工程师实际工作方式的一项架构要求。

从战略角度看，这正是整个组织所需要的能力：在任意时刻自由使用最好的可用工具，同时不必在生态系统每次演进时都承担迁移成本。Xirp 让工程师能够在任务中途切换模型，并将每项工作路由至价格性能最优的可用选项，包括由我们自行托管的开源模型，从而确保我们不会被锁定。

当新模型出现时，我们的工程师无需重建自己的环境就能完成迁移。他们能够在不同编程智能体之间流畅切换、跨会话保留上下文，并编排并行工作流。

我们在内部看到了良好的验证：数千名 Spotify 工程师已在超过 36,000 个会话中自然采用 Xirp，并实现了更快的上下文切换和成本效率提升。

Xirp 本身解决了多智能体编排挑战；当它与 Portal 配合使用时，便成为真正的智能体指挥中心。

当 Xirp 连接到 Portal 时，每个智能体会话都会使用从软件目录中提取的组织上下文完成初始化：组件架构、依赖关系图、所有权拓扑，以及规定系统如何相互关联的架构决策。

结果是，整个工程组织都能基于同一个共享的机构知识基线开展工作。曾使 Backstage 对人类工程师产生变革性影响的同一套结构化知识，如今也让 AI 智能体的效率显著提升。

这种价值会向两个方向叠加。每次会话结束后，记录文本和元数据都会回流至 Portal，为整个组织提供全面可见性：已经完成了什么、谁在处理什么，以及应当从哪里继续。任何工程师或智能体都可以接手另一方未完成的工作，因为完整的工作上下文都保留在 Portal 中。

团队内的工作变得可互换。重复投入被消除。一名工程师的智能体会话所产生的组织知识，会立即可供团队中后续的每一个会话使用。

考虑一下每个团队独立开发的技能、规则、插件和 MCP 配置。Portal 将它们整合到一个经过策展的市场中，团队可以在其中贡献、发现并在彼此的工作基础上继续构建。对于偏好直接在所选智能体中工作的团队，Portal 的 MCP 工具能够在任何编程会话中原生呈现同样的组织上下文。

Spotify 的工程文化建立在这样一种信念之上：赋予团队快速行动的自主权，但同时投资于能够让这种速度可持续的共享基础设施。我们将这一原则用于 Backstage 和 Portal 的开发者体验建设，用于 Confidence——我们的实验平台——的产品开发，如今也将其用于借助 Xirp 和 Portal 推进智能体软件开发。

那些学习速度最快、能够将工程知识沉淀为共享且结构化的系统，而不是让其散落在个人手中的组织，将会是获得 AI 辅助开发全部价值的组织。我们已经在大规模实践中证明了这一模式。现在，我们正将这份专业能力带给你。

Xirp 今天即可作为独立的智能体开发环境创造价值。将它连接到 Portal，组织上下文就会成为放大器，持续提升每一个智能体、每一次会话和每一个团队的效率。

如需了解更多信息或免费试用，请访问 xirp.spotify.com。

Every CTO I speak with has moved past the adoption question. Their engineering organizations are using AI coding agents daily. But the landscape beneath them is in constant motion with new models, new agents, rising costs, pricing structures that shift quarter to quarter, and capabilities that advance overnight.

The real challenge is no longer velocity alone. It is ensuring teams have the flexibility to adopt whatever maximizes value at any given moment, without accumulating vendor dependency or compounding costs as you scale.

At Spotify, with thousands of engineers operating across hundreds of teams, this tension defines how we approach our development infrastructure. Our engineers require the autonomy to adopt the best available tool for their work, and what qualifies as best evolves rapidly. But autonomy without structure produces fragmentation, duplicated effort, and inconsistent outcomes. Reconciling that tension is what has driven our approach.

The progression was predictable in hindsight. Teams began with a single agent. One session, one developer, one task. Then adoption accelerated.

Engineers, and later on non-engineers, started orchestrating multiple agents in parallel, across different sessions and tools, repositories, and branches. Context accumulated in one session but was needed in another. Institutional knowledge fragmented into individual configurations — CLAUDE.md files, bespoke MCP setups, personal prompt libraries — making it harder to carry that context across teams and workflows.

The resulting entropy impacted productivity. PRs increased, but so did rework, inconsistency, and wasted tokens. Agents in one session would expend cycles rediscovering what another session had already resolved.

Moreover, engineers were investing as much effort reconstructing context as they were on building new capabilities. And at the organizational level, creating greater visibility across thousands of concurrent sessions became an increasingly important priority.

This was not a new class of problem for us. A decade ago, the complexity of thousands of engineers operating across thousands of services produced the same category of fragmentation. That is why we built Backstage, a unified, structured source of truth for how our systems interconnect: dependency maps, component metadata, ownership, architectural decisions. Backstage became the foundation of our internal developer platform, and we subsequently built Portal, so any organization can get the same benefits without managing the infrastructure themselves.

AI coding agents have introduced an entirely new dimension of that same organizational complexity. It requires its own solution.

Xirp is a vendor-neutral agentic development environment born out of a concrete engineering need.

Xirp was built to help our developers manage dozens of concurrent agent sessions across multiple harnesses (Claude Code, Gemini CLI, Codex, etc.), making it tenable to coordinate 50+ parallel sessions. Every session operates in its own worktree, enabling dozens of agents to work concurrently on the same codebase without interference. Context is decoupled from any single agent or harness; switch tools mid-project, and the full working state carries over. Vendor neutrality is an architectural requirement that emerged directly from how our engineers work.

From a strategic perspective, this is precisely the capability needed across our organization: the freedom to leverage the best available tool at any given moment, without incurring a migration cost every time the ecosystem evolves. Xirp lets our engineers switch models mid-task and route every job to the best available price performance, including open source models that we host ourselves, so that we are never locked in.

When a new model emerges, our engineers transition without reconstructing their environment. They are able to move fluidly between coding agents, preserve context across sessions, and orchestrate parallel workstreams.

We're seeing good validation internally: thousands of Spotify engineers have organically adopted Xirp across more than 36,000 sessions, resulting in faster context switching and cost efficiencies.

Xirp on its own addresses the multi-agent orchestration challenge, and when paired with Portal, it becomes a true agent command center.

When Xirp connects to Portal, every agent session initializes with organizational context drawn from your software catalog: component architecture, dependency graphs, ownership topology, and the architectural decisions that govern how systems interrelate.

The result is a shared institutional baseline your entire engineering organization operates from. The same structured knowledge that made Backstage transformative for human engineers now makes AI agents dramatically more effective.

The value compounds in both directions. After each session, transcripts and metadata flow back into Portal, providing comprehensive visibility across the organization: what has been accomplished, who is working on what, and where to resume. Any engineer or agent can pick up where another left off, because the full working context is preserved in Portal.

Work becomes interchangeable across the team. Duplicate effort is eliminated. The institutional knowledge that one engineer's agent session produces becomes immediately available to every subsequent session across the team.

Consider the skills, rules, plugins, and MCP configurations that every team develops independently. Portal consolidates it into a curated marketplace where teams contribute, discover, and build upon each other's work. For teams that prefer to operate directly within their agent of choice, Portal's MCP tools surface the same organizational context natively within any coding session.

Spotify's engineering culture was built on a conviction: give teams the autonomy to move fast, but invest in the shared infrastructure that makes that speed sustainable. We applied that principle to developer experience with Backstage and Portal, to product development with Confidence, our experimentation platform, and now to agentic software development with Xirp and Portal.

The organizations that learn fastest, that compound their engineering knowledge into shared, structured systems rather than leaving it scattered across individuals, will be the ones that capture the full value of AI-assisted development. We have proven that pattern at scale. Now we're shipping that expertise to you.

Xirp delivers value as a standalone agentic development environment today. Connect it to Portal, and organizational context becomes the multiplier, compounding the effectiveness of every agent, every session, every team.

To learn more or try it for free, visit xirp.spotify.com.