中文精读
给Agent造环境：价值在反馈、边界和中断后还能继续
42章经单期精读 · Agent基础设施
嘉宾：雷磊，Grasp、AirCode创始人，前字节开发者平台轻服务产品负责人，身份依当期介绍｜主持：曲凯
发布：2025年6月14日｜时长：54分35秒｜整理：2026年10月18日
官方节目与收听入口
当软件的使用者变成Agent
一个自动办事的Agent，需要的不只是聪明模型。它要运行代码、浏览网页、保持登录、等待人确认，再回到先前的位置。雷磊把为这些行为提供的环境称为Agent Infra，即Agent基础设施。他希望Grasp承担浏览器的通用工程，让应用开发者集中处理行业任务。
本期的核心问题是：哪些工程问题真的因Agent而变得重要，怎样证明它们值得付费？2025年节目发布时，官方页仍称Grasp开放等待名单、即将内测；访谈中还提到预计下个月开放。功能、安全与竞争优势因此按当时团队陈述理解，不能认作当前已验收结果。
1. 反馈要来自任务结果，不能只看动作像不像
节目线索：01:41–12:41。 雷磊把SaaS理解为人使用的工具，把Agent理解为交付结果的主体。通用与垂直Agent可像商场和小店共存：功能覆盖面不同，能服务某一群人的具体结果才是关键。
他用写代码解释新的环境需求。Agent可以同时尝试多个方法，运行后获得反馈，再选择下一步。因此环境要低成本启动、快速返回结果，并保留失败原因。访谈用“人单线程、AI多线程”及局部最优、全局最优作类比，但这不是人只能串行思考、Agent保证全局最优的技术结论。
更重要的区分是grounded signal，即来自实际行动结果的信号。浏览器点击顺序看起来合理，订单却未提交，就不能算完成。雷磊希望把执行结果回馈系统，但节目未展示训练效果。编辑推论：验收应同时保留最终回读和任务轨迹：结果决定是否完成，轨迹帮助解释失败、责任和是否发生多余动作。
2. 沙盒先回答：最坏能影响到哪里
节目线索：13:03–20:51、24:16–26:05。 AI生成代码可能误删文件，运行环境需要围栏。沙盒是限制影响的执行空间；虚拟化是一类实现技术；云是一种资源交付方式；E2B则是把隔离、启动与使用接口组合起来的产品。它们不是同义词。
雷磊强调围栏大小的取舍：太小，任务做不成；太大，损失不可接受。他介绍E2B借MicroVM缩短启动时间，但没有同条件性能测试，不能把“MicroVM一定比容器快”当选型原则。安全也不是买一个名词：允许读写什么文件、访问什么网络、能否影响外部账户，都要具体说明。
雷磊认为Agent环境会云端化，理由是弹性、持续运行与靠近模型。这里保留他的判断；本地同样可以隔离和持续服务，最终选择还取决于延迟、数据、控制和运维。云厂商与环境厂商被比成资源商和装修商，帮助理解分工，却不能保证未来没有竞争。
自建还是购买也由需求决定。他倾向使用成熟方案，只有关键需求未被满足时才自建或定制。编辑推论：先写任务必须具备的权限、恢复与成本条件，再比较供应商，代码归谁写不是价值的充分证据。
3. 中断一小时再回来：状态与资源要分开保存
节目线索：21:08–24:48、34:03–35:36。 ForeverVM的例子把痛点讲得很具体：Agent运行一段脚本，去做别的事，一小时后还要接着运行。如果环境一直开着，会付出闲置资源成本；如果直接销毁，前面的状态可能丢失。保留进度、暂停资源消耗，成为一类产品价值。
浏览器也一样。Agent查完机票，把信息交给另一系统推理，可能等待人决定，之后再回原页面。雷磊介绍有状态浏览器会话，希望用户感觉网页始终在那里，期间却不用持续占用相同资源。
这里的“还在”需要拆开：页面位置、登录、已输入数据、服务端价格及库存未必同时有效。主持人也提到普通人返回购票页面会遇到价格更新。编辑推论：恢复成功应检查任务能否继续，而非只看截图相似；真实价格与权限要重新确认。节目没有提供节省成本比例、恢复成功率或状态保留期限，不能替团队补上承诺。
4. 给Agent的浏览器：接口、凭据与三层责任
节目线索：31:54–40:20。 Headless即后台运行、没有供人操作的前端窗口。Agent未必要像人那样逐个用鼠标点页面，可以读取结构化信息、执行脚本；但这不意味着所有任务永远不需要视觉，也不意味着人不再需要可见界面。
凭据是另一个难点：既希望Agent自主登录，又不想把密码交给模型。雷磊介绍Grasp的安全本地登录，声称能以本地方式填账号、密码或验证码，避免传入大模型。这里是产品方的安全主张，本期没有流量、日志、权限与故障测试可验证，不能写成已经证明的绝对保证。
他把浏览器Agent拆成三层：Runtime运行时处理页面获取、脚本和渲染；Agentic层处理观察、推理与行动；Knowledge层提供行业知识、反馈和验收。Grasp当时希望整合下面两层，让开发者带着行业理解做应用。与Browserbase和Stagehand的比较，是2025年的供应方视角，不能直接当当前产品能力清单。
三层划分对产品经理的用处是分责任：网页没加载、模型选错按钮、订单业务条件判断错，可能需要不同团队处理。接口封装减少工程负担，但应用方仍要定义最终结果和授权范围，不能购买一个浏览器就认为行业Agent完成了。
5. 面向未来的市场，入口仍是今天的旧流程
节目线索：28:56–31:29、41:06–47:44、51:23–54:10。 雷磊介绍Browserbase时，先讲已有爬取、自动化测试与RPA（按规则自动操作软件）需求。网页变化、网络延迟与状态管理本来就会让自动化失败；Agent可能提供新的解决方式，而非凭空创造所有需求。
节目中的“40%流量来自机器”、估值3亿美元、软件市场数万亿美元等，均为当期口述且未核方法。以市场规模乘5%效率推算机会，也不能直接得到基础设施厂商收入：省下的价值如何收费、谁愿意付、维护成本多少，还要验证。
雷磊认为代码与浏览器是两类重要环境，前者执行逻辑，后者与Web信息及行动交互；数学、空间与传感器也可能产生其他环境。他没有给出Agent全面普及的确定时间表，最终仍回到营销、销售、测试等已有用户。行业增长也不意味着某家旧公司的股票必涨，访谈对此有明确保留，不构成投资建议。
6. “让AI自己体验”，仍需要人定义什么是好结果
节目线索：45:45–51:09。 嘉宾用菜谱作例子：只请厨师看文字评价，可能限制在已有经验；把菜真正做出来再尝，可以获得行动后的反馈，发现文字直觉未预测到的结果。他以此强调环境对学习的重要性。
这是一种研究方向，不能扩成“人类反馈毫无价值”。好吃由谁判断、成本与安全是否可接受、是否符合使用者目标，仍需要定义。可验证结果也不都像数学答案一样明确。编辑推论：业务团队应先区分机器可直接检查的结果、人必须判断的体验、以及无权自动尝试的动作，再设计奖励；否则系统可能优化一个漂亮却错误的代理指标。
用一次中断任务比较环境方案
以下是编辑建议。选一个“查找信息—等待确认—回原页面继续”的任务，让两个方案完成同样输入，记录：
| 环节 | 必须看到的证据 | 不能替代的信号 |
|---|---|---|
| 执行 | 真实结果回读、失败原因 | 动作动画顺畅 |
| 边界 | 文件、网络、账户允许范围与实际日志 | 笼统声称安全 |
| 等待恢复 | 暂停成本、恢复耗时、登录和数据是否有效 | 同一张页面截图 |
| 业务结果 | 条件复核、授权与最终回执 | 工作流正常退出 |
先用现存任务验证，再决定是否增加并行和投入。“为Agent重做软件”的机会，应从一个具体失败被稳定解决开始。
回听导航与关联阅读
- 10:49：基础设施为什么要重视反馈。
- 13:03：E2B与沙盒边界。
- 21:02：深度使用者如何发现状态问题。
- 31:48：AI浏览器的交互与登录。
- 35:55：Runtime、Agentic、Knowledge三层。
- 49:04：菜谱与真实世界反馈。
相关主题：Agent架构、执行环境、质量评估。可对照Dify早期工程与持续运营和Albert对评估的强调。
来源边界：完整阅读提供转写、旧精读和官方当期介绍后重新组织，未逐句复听。名称按官方资料与清晰上下文校正。性能、安全登录、竞品能力、估值、规模与市场推演均归2025年访谈陈述，未另做技术、安全或财务验证；产品计划不代表当前状态。
English edition
Building Environments for Agents: Feedback, Boundaries and Resuming After an Interruption
42 Zhangjing episode analysis · Agent infrastructure
Guest: Lei Lei, founder of Grasp and AirCode and formerly responsible for lightweight service products on ByteDance’s developer platform, as introduced at the time | Host: Qu Kai
Published: June 14, 2025 | Duration: 54 minutes 35 seconds | Prepared: October 18, 2026
Official episode and listening link
When software’s user becomes an agent
An agent performing tasks needs more than a capable model. It must run code, browse websites, preserve login, wait for confirmation and return to its previous position. Lei calls the supporting environments Agent Infra, or agent infrastructure. He wants Grasp to handle common browser engineering so application developers can concentrate on industry tasks.
Which engineering problems become more important with agents, and what demonstrates that solving them merits payment? At publication in 2025, the official page described a waiting list and an approaching beta; the conversation also mentions opening the following month. Features, security and competitive advantages therefore remain contemporary team statements, not currently verified delivery.
1. Feedback should reflect completed work, beyond plausible-looking actions
Episode reference: 01:41–12:41. Lei frames SaaS as a tool people use and an agent as an entity delivering results. General and specialized agents could coexist like shopping centers and small shops. Different coverage matters less than serving a group through concrete outcomes.
He uses coding to explain environment requirements. Agents can try multiple methods, execute them, obtain feedback and select their next steps. Environments should consequently start cheaply, return results quickly and retain failure causes. The interview’s comparisons between single-threaded people and parallel AI, or local and global optimization, are teaching analogies. They do not establish that people only think serially or that agents guarantee global optima.
The more useful distinction is a grounded signal: feedback from an actual outcome. A plausible sequence of clicks does not complete an unsubmitted order. Lei wants execution results fed back into the system, but the episode demonstrates no resulting training improvement. Editorial inference: Retain both final readback and the task trace. Outcomes establish completion; traces help explain failures, responsibility and unnecessary actions.
2. A sandbox should explain the largest possible impact
Episode reference: 13:03–20:51 and 24:16–26:05. Generated code might delete files, requiring a fence around execution. A sandbox limits effects; virtualization supplies implementation techniques; cloud describes resource delivery; E2B combines isolation, startup and interfaces into a product. These are different concepts.
Lei emphasizes the fence’s size. Too narrow and tasks fail; too broad and losses become unacceptable. He describes E2B using MicroVM technology to accelerate startup, without comparable performance tests. “MicroVMs are always faster than containers” cannot become a selection rule. Security also needs specifics: which files, networks and external accounts can execution affect?
He expects agent environments to move to cloud infrastructure because of elasticity, persistent operation and proximity to models. This remains his judgment. Local systems can also provide isolation and continuous services; latency, data, control and operations affect the choice. Resource providers and interior contractors illustrate the relationship between cloud and environment companies, without guaranteeing an absence of future competition.
Building versus buying similarly depends on requirements. Lei favors mature solutions unless a critical unmet need requires custom development. Editorial inference: Specify permissions, recovery and cost conditions first, then compare suppliers. Authorship of code alone does not establish value.
3. Return an hour later: Preserve state separately from resource consumption
Episode reference: 21:08–24:48 and 34:03–35:36. ForeverVM illustrates an agent executing a script, moving to another activity and returning an hour later. Keeping the environment running incurs idle costs; destroying it risks losing state. Preserving progress while pausing resource use creates a product opportunity.
A browser has the same problem. An agent checks flights, sends information elsewhere for reasoning, waits for a human decision and then returns. Lei describes stateful browser sessions intended to feel continuously available without maintaining the same resource allocation throughout the wait.
“Still there” needs decomposition. Page position, login, entered data, server-side price and inventory may have different validity. The host notes that people also encounter updated fares on returning. Editorial inference: Successful recovery means the task can continue, beyond a similar screenshot. Prices and permissions need rechecking. No cost-saving ratio, recovery success rate or retention duration is supplied; these cannot be added as product promises.
4. Browsers for agents: Interfaces, credentials and three layers of responsibility
Episode reference: 31:54–40:20. Headless means running in the background without a front-end window for human operation. An agent may read structured information or execute scripts instead of clicking with a mouse. This does not establish that vision is never needed or that people no longer need a visible interface.
Credentials create another tension: autonomous login without giving passwords to the model. Lei describes Grasp’s secure local login, claiming local entry of credentials and verification codes without transmitting them to a large model. This is the supplier’s security claim. Traffic, logs, permissions and failure tests are absent, so it cannot become an independently proven absolute guarantee.
He separates browser agents into three layers. Runtime fetches pages, executes scripts and renders. The Agentic layer observes, reasons and acts. Knowledge supplies domain expertise, feedback and acceptance. Grasp proposed integrating the lower two layers so developers could bring industry understanding. Comparisons with Browserbase and Stagehand reflect a supplier’s 2025 perspective, not a current capability inventory.
The framework helps product managers assign responsibility. A page failing to load, a model choosing the wrong button and a mistaken order condition may require different teams. Encapsulation can reduce engineering work, but application owners must still define outcomes and authorization. Buying a browser does not complete an industry agent.
5. A future market still begins with existing workflows
Episode reference: 28:56–31:29, 41:06–47:44 and 51:23–54:10. Lei introduces Browserbase through existing scraping, testing and RPA, or rule-based software automation. Changing pages, network delays and state management already break automation. Agents may offer new solutions rather than create every need from nothing.
Machine traffic at 40%, a US$300 million valuation and software markets worth trillions are contemporary spoken claims with unverified methods. Multiplying a market by a 5% efficiency gain also does not directly produce infrastructure revenue. Pricing, willingness to pay and maintenance costs remain separate questions.
Lei treats code and browsers as important environment categories: executing logic and interacting with web information and actions. Mathematics, space and sensors may create others. He offers no definite schedule for widespread adoption, returning to existing marketing, sales and testing customers. Industry growth also does not guarantee gains in a particular incumbent’s shares; the discussion explicitly preserves this distinction and is not investment advice.
6. Learning from experience still requires defining a good outcome
Episode reference: 45:45–51:09. Lei compares asking a chef to judge a recipe with cooking and tasting it. Textual judgments may remain within existing experience, while action can reveal outcomes that intuition failed to predict. He uses this to emphasize environments in learning.
This research direction does not establish that human feedback has no value. Someone must still define taste, acceptable cost, safety and user goals. Not every outcome is as unambiguous as a mathematical answer. Editorial inference: Separate machine-checkable results, human-judged experiences and actions that cannot be attempted without authorization before defining rewards. Otherwise a system may optimize an appealing but mistaken proxy.
Compare environments through one interrupted task
This is an editorial exercise. Give two solutions the same information-search, confirmation-wait and browser-return task, then record:
| Stage | Required evidence | Insufficient substitute |
|---|---|---|
| Execution | Final readback and failure causes | Smooth action animation |
| Boundary | Allowed files, networks and accounts, plus actual logs | A general security claim |
| Recovery | Idle cost, recovery time, valid login and data | The same page screenshot |
| Business outcome | Rechecked conditions, authorization and final receipt | A workflow exiting normally |
Verify an existing task before expanding parallelism and spending. The opportunity to rebuild software for agents begins with reliably fixing a specific failure.
Listening guide and related reading
- 10:49: Why infrastructure needs feedback.
- 13:03: E2B and sandbox boundaries.
- 21:02: Intensive users discovering state problems.
- 31:48: Browser interaction and login.
- 35:55: Runtime, Agentic and Knowledge.
- 49:04: Recipes and real-world feedback.
Related themes: agent architecture, execution environments and evaluation. Compare early Dify’s engineering and operations and Albert’s emphasis on evaluation.
Source boundary: Reorganized after complete reading of the supplied transcript, prior analysis and official introduction, without sentence-by-sentence audio checking. Names follow official material and clear context. Performance, secure login, competitors, valuations, scale and market projections remain attributed to the 2025 conversation, without separate technical, security or financial validation. Plans do not establish current product status.