Warp(在新窗口中打开) 最初是一款现代终端,凭借其速度、协作功能、命令工作流和 AI 原生界面,早期就赢得了开发者的青睐。随着编码智能体从实验阶段进入日常工程工作流,Warp 发现终端正成为开发者与智能体协作的自然场所:命令、上下文、协作和审查在此交汇。
今年,当 Warp 开源(在新窗口中打开) 其终端客户端时,OpenAI 作为该仓库的创始赞助商,公司还推出了开放智能体开发模式:一种在开放环境中构建软件的模式。人类定义目标并监督结果,而智能体则负责规划工作、编写代码、测试变更并提交拉取请求。
前沿 AI 模型的最新进展帮助这种智能体编排在大规模场景下变得实用。对于 Warp 的开源工作流,GPT‑5.5 帮助智能体在更大的问题空间中进行推理,并为人工审查准备工作。在内部基准测试中,GPT‑5.5 在每个智能体编码任务中使用的 token 比 GPT‑5.4 少 30%,帮助 Warp 在扩展长期运行的智能体工作流时提高效率。
如今,Warp 拥有近 100 万开发者,并被超过 56% 的财富 500 强企业使用。在 Warp 自己的工程团队中,智能体现已共同创建了公司约 90% 的拉取请求,这让团队对长期运行的智能体工作流扩展所需的条件有了第一手了解:可观测性、协调、记忆和人工审查。
“我们认为,通过与社区合作监督一组智能体,我们可以更快地交付更好的 Warp。OpenAI 模型有助于使这些系统所需的长期编码工作变得可持续。”
—Zach Lloyd,首席执行官
下一代协作式软件开发
开放智能体开发是 Warp 对软件开发未来方向的押注。智能体将编写代码,而开发者将指定意图、验证输出,并决定最终发布的内容。这些选择将成为未来智能体的可复用上下文,使系统能够随着时间的推移不断改进。
Warp 认为,如果编排足够出色,智能体可以比松散协调的人类团队生成更一致的代码。开源将不再侧重于人类直接贡献实现工作,而是更多地贡献只有人类才能提供的产品判断和共同愿景。
持久化、并行化的智能体需要共享内存、可重现环境、评估系统、权限以及协调工作的方法等组件。Warp 构建了其云编排平台 Oz,用于在本地和云环境中管理智能体。
据公司称,对于开放智能体开发工作流,Warp 使用 GPT‑5.5 来管理帮助处理开源贡献的智能体。OpenAI 模型在 Warp 内部针对涉及推理、规划、代码生成和代码审查的长期工程任务的评估中也表现出色。
使用 Oz 进行智能体编排
Oz(在新窗口中打开) 充当控制平面,用于在本地和云环境中部署和协调智能体。开发者可以通过 Web 界面启动智能体,选择预定义的技能和环境,选择模型和托管配置,并在执行过程中集中监控长期运行的工作流。
一旦启动,智能体可以继续远程运行,同时开发者可以检查实时会话、监控执行状态、审查生成的工件,并在云环境和本地环境之间来回切换工作流,而不会丢失上下文。Oz 还支持定期工作流,使智能体能够像计划中的 cron 作业一样运行。
随着智能体随时间积累更多状态,保持专注并保留重要决策变得越来越困难。Oz 使用上下文压缩、持久化内存以及用于代码搜索和文件分析等任务的专用子智能体等技术,帮助智能体在扩展工作流中保持可靠性。
OpenAI 模型在 Oz 中扮演多个角色。对于 Warp 智能体,任务按类型和难度进行分类,更复杂的编码和推理工作被路由到更强的模型配置。GPT‑5.5 是 Warp 用于要求苛刻的智能体编码工作流的 OpenAI 模型组合的一部分。Warp 还在其评估管道中使用 OpenAI 模型作为 LLM 作为评判系统。
“我们发现,OpenAI 模型在完成相同任务时,通常能提供前沿水平的智能,同时使用更少的 token 和更少的轮次。这些模型在需要跨大问题空间进行推理的编码任务中尤其强大。”
—Zach Lloyd,首席执行官
构建智能体开发的基础设施
对于 Warp 来说,开放智能体开发和 Oz 编排平台最终都是同一个长期押注的一部分:软件开发正从与编码助手的个人交互,演变为协调大量持久化智能体的系统。
到目前为止,这个押注似乎正在获得回报。Warp 的 ARR 去年增长了 35 倍,自 2025 年第四季度以来,企业收入增长了超过 500%。公司表示,这一增长主要来自那些寻求更灵活方式来扩展智能体工作流的组织。
围绕智能体开发的底层工作流仍处于早期阶段,且高度实验性。通过开源其终端客户端,并利用开放智能体开发工作流公开构建,Warp 希望开发者能够帮助塑造编排、监督和验证系统在智能体日益自主化过程中的演变方式。
“没有人确切知道智能体开发的未来会是什么样子,”Lloyd 说。“我们认为社区应该能够参与塑造它。”
Warp(opens in a new window) started as a modern terminal, earning early love from developers for its speed, collaboration features, command workflows, and AI-native interface. As coding agents moved from experiments to everyday engineering workflows, Warp saw the terminal becoming a natural place for developers to work with agents: where commands, context, collaboration, and review already meet.
When Warp open-sourced(opens in a new window) its terminal client this year, with OpenAI as the founding sponsor of the repo, the company also introduced Open Agentic Development: a model for building software in the open. Humans define objectives and supervise outcomes, while agents plan work, write code, test changes, and open pull requests.
Recent improvements in frontier AI models helped make that kind of agent orchestration practical at scale. For Warp’s open-source workflows, GPT‑5.5 helps agents reason across larger problem spaces and prepare work for human review. In internal benchmarks, GPT‑5.5 used 30% fewer tokens per agentic coding task than GPT‑5.4, helping Warp improve efficiency as it scales long-running agent workflows.
Today, Warp has nearly 1 million developers and is used by more than 56% of the Fortune 500. In Warp’s own engineering organization, agents now co-create around 90% of the company’s pull requests, giving the team a firsthand view into what long-running agent workflows need to scale: observability, coordination, memory, and human review.
“We think we can ship a better Warp, more quickly, by working with our community to supervise a fleet of agents. OpenAI models help make that sustainable for the long-horizon coding work these systems require.”
—Zach Lloyd, CEO
The next generation of collaborative software development
Open Agentic Development is Warp’s bet on where software development is heading. Agents will write code, and developers will specify intent, verify the outputs, and decide what ultimately ships. Those choices become reusable context for future agents, allowing the system to improve over time.
If the orchestration is good enough, Warp believes agents can produce more consistent code than a loosely coordinated group of humans. Open source then becomes less about humans contributing implementation work directly, and more about contributing the product judgment and shared vision that only humans can provide.
Persistent, parallelized agents need components like shared memory, reproducible environments, evaluation systems, permissions, and ways to coordinate work. Warp built Oz, its cloud orchestration platform, to manage agents across local and cloud environments.
For Open Agentic Development workflows, Warp uses GPT‑5.5 for agents that help manage open-source contributions, according to the company. OpenAI models have also performed strongly in Warp’s internal evaluations for long-horizon engineering tasks involving reasoning, planning, code generation, and code review.
Agentic orchestration with Oz
Oz(opens in a new window) acts as a control plane for deploying and coordinating agents across local and cloud environments. Developers can launch agents through a web interface, select predefined skills and environments, choose the model and hosting configurations, and monitor long-running workflows centrally as they execute.
Once launched, agents can continue running remotely while developers inspect live sessions, monitor execution state, review generated artifacts, and hand workflows back and forth between cloud and local environments without losing context. Oz also supports recurring workflows, allowing agents to operate like scheduled cron jobs.
As agents accumulate more state over time, maintaining focus and preserving important decisions becomes increasingly difficult. Oz uses techniques like context compaction, persistent memory, and dedicated subagents for tasks like code search and file analysis to help agents stay reliable across extended workflows.
OpenAI models play several roles inside Oz. For the Warp agent, tasks are classified by type and difficulty, with more complex coding and reasoning work routed to stronger model configurations. GPT‑5.5 is part of the OpenAI model mix Warp uses for demanding agentic coding workflows. Warp also uses OpenAI models as LLM-as-a-judge systems inside its evaluation pipelines.
“We’ve found that OpenAI models regularly provide frontier-level intelligence while taking fewer tokens and turns to complete the same tasks. The models are especially strong for coding tasks that require reasoning across large problem spaces.”
—Zach Lloyd, CEO
Building the infrastructure for agentic development
For Warp, Open Agentic Development and the Oz orchestration platform are ultimately part of the same long-term bet: that software development is evolving from individual interactions with coding assistants into systems for coordinating large numbers of persistent agents over time.
So far, that bet seems to be paying off. Warp’s ARR grew 35x last year, with enterprise revenue up more than 500% since Q4 2025. The company says much of that growth is coming from organizations looking for more flexible ways to scale agent workflows.
The underlying workflows around agentic development are still early and highly experimental. By open sourcing its terminal client and building in public with Open Agentic Development workflows, Warp hopes developers can help shape how orchestration, supervision, and verification systems evolve as agents become more autonomous over time.
“No one knows exactly what the future of agentic development will look like,” Lloyd says. “We think the community ought to be able to participate in shaping it.”
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