QoderWake

• 1. overview (579 chars)

  QoderWake is a "digital employee" product released by Alibaba's Qoder team on April 30, 2026. It is positioned as a 24/7 on-the-job colleague that is production-available, safe and controllable, and capable of self-evolution. The fundamental difference between it and traditional AI programming assistants lies in the collaboration logic: from "people looking for AI" to "AI looking for people" - taking over autonomously after an event is triggered, reporting after completing the work, and people only confirm at key nodes. The first "Digital Programmer" role has been deeply used within Alibaba, and the root cause analysis time has been reduced from about 30 minutes to 2 minutes.

• 2. Product background (845 chars)

  Qoder is Alibaba's real-work-oriented agent platform. It competes with Cursor and Windsurf for the same type of development workflow. It relies on the Tongyi series of large models. After the independent IDE was launched on August 22, 2025, Teams team version, QoderWork office agent, and QoderWake digital employee were launched intensively within half a year. According to official disclosures, Qoder’s entire range of products currently has more than 5 million users, and its ARR exceeded US$60 million, only 8 months after its launch.   QoderWake directly benchmarks Devin and traditional RPA: Devin is an "execution type" and can complete complex tasks but lacks the security constraints of a production environment; QoderWake is one of the few products that is clearly positioned as an "employee type". It has job identity, self-evolving memory, strict authority control and around-the-clock duty capabilities, and is specially designed to integrate into enterprise production processes.

• 3. Product functions and usage experience (1559 chars)

  The core of QoderWake is a Harness-First architecture. After the task is executed, the system will automatically classify the experience into the five dimensions of memory, skills, strategies, verification rules and workflow, solving the problem of "forgetting once it is done" for general-purpose Agents; the built-in anti-corrosion mechanism will continue to eliminate outdated experience, merge conflicting skills, and withdraw invalid strategies to ensure that the more you use it, the more accurate it will be.   The first batch of public beta opens six built-in positions that have been systematically trained: back-end engineer, front-end development engineer, test engineer, product manager, data analyst, content operations specialist, and supports fully customizable roles (setting responsibilities, work styles, tool permissions). The skills level has built-in more than 100 job skills, and is equipped with a skills market of hundreds of selected skills, covering various scenarios, and the capabilities can be continuously expanded with demand.   At the access level, digital employees can connect with GitHub, Jira, and GitLab, support DingTalk, and Feishu, and automatically start work based on time or events. It is suitable for handling repetitive work that occurs repeatedly, has clear boundaries, and can accept the results. Several internal beta users mentioned that they configured a "front-end assistant" to handle component restoration tasks. It originally took half a day and 20 minutes to produce the first draft, and also automatically ran Typecheck and basic tests. Some people also "recruited" digital content operation employees to analyze the recommendation flow and batch generate Xiaohongshu information flow diagrams. After installing more than 20 Skills, their capabilities are quite solid.

• 4. Pricing and business model (469 chars)

  The pricing of QoderWake itself has not been disclosed yet and is currently in the public beta stage. The Qoder system to which it belongs is billed by credits: the Pro version of IDE is about US$30/month (2000 credits), Pro+ is US$60/month, Ultra is US$200/month, and Teams is US$40/seat/month. QoderWake supports access to domestic and foreign mainstream models. You can switch between GLM-5.1, DeepSeek V4 Pro, Kimi-K2.6, MiniMax-M2.7, Qianwen 3.7-Max, or use auto mode to intelligently select models based on tasks. Behind Ultimate is Claude Opus 4.7.

• 5. User reviews (1244 chars)

  Positive feedback focused on "collaboration feels like colleagues" and "worry-free". Many users said that digital employees have identities, memories, and experience, and task results are automatically accumulated. Next time they encounter similar needs, they can refer to previous handling methods; work records are traceable, and it is clear at a glance who did it, when, and what verifications were run. After connecting to DingTalk, it can dispatch tasks, and the response speed is faster than waiting for a real person to reply. Some users also compare QoderWake with QoderWork, saying that the former is an "unblinking night watchman". If there is an online problem in the middle of the night, people no longer need to get up to read the log.   In terms of negatives and shortcomings, internal beta users generally mentioned two points: First, "delivery is not intuitive" after running the task. The results are placed in local folders, but are not displayed on the page. The generated file cannot be found if the workspace is not specified at the beginning, so it must be given an absolute path automatically; second, they hope to add the dashboard and task management list, and assign tasks to different employees according to priority and deadline to automatically claim them. In addition, the current public beta only provides the Mac version. Some positions (such as AI finance) have not yet been connected to the actual business system, and what can be done is still limited.

• 6. Industry evaluation (589 chars)

  Industry media regard Qoder as a powerful substitute for Cursor in heavy-load workloads of agents, affirming the ability of Quest's autonomous mode to deliver complete functions at once. In terms of technical positioning, the difference between QoderWake and Devin has been repeatedly emphasized - the former is the only "employee" Agent that takes production-level security constraints as the core of the product. Some analysts believe that Alibaba’s move is very clear: QoderWork competes for desktop access and competes with Claude Cowork, while QoderWake competes with enterprise positions and competes with Devin and traditional RPA. The two are complementary rather than substitutes.

• 7. Controversies and Risks (730 chars)

  QoderWake's trust boundary is based on permission redlines and approval mechanisms: destructive commands are blocked by default, files are whitelisted for reading and writing by directory and automatically backed up, high-risk operations are stopped and asked for permission first, and each task has a complete execution record for review. But risks also exist - it is still in public beta, and the boundaries of capabilities are far from closed; "autonomous work" means that once the authority is released too wide, the impact of misoperation will be magnified. In addition, the so-called "memory accumulation, the more you use it, the more accurate it becomes" relies on long-term running data. The initial experience of new users may not be as "sensible" as advertised. They need to run a few tasks with read-only permissions before gradually opening up writing.

• 8. Applicable groups and usage suggestions (539 chars)

  It is most suitable for teams that have on-call requirements and need to be on duty: online operation and maintenance, code review, alarm processing, root cause analysis, and free people from "tracking" repetitive work. If individual developers only want to improve daily office efficiency, the free version of QoderWork is lighter and has a lower threshold; if both daily efficiency and operation and maintenance are required, the two can be combined. It is recommended to run 1-2 tasks with "read-only" permission when configuring for the first time, and then gradually open write permission after confirming that the logic is reliable.

• 9. Summarize (345 chars)

  QoderWake upgrades AI from "short-term tool" to "long-term colleague". Its essence is to use identity, memory and governance mechanisms to solve the trust problem of general-purpose Agent production. It may not be suitable for every job for everyone, but for repetitive tasks with clear boundaries and on-duty tasks, it has already shown the prototype of a digital employee that people are willing to "hire".

No in-depth report yet

• 

  Sharon.MorganSr

  —

  新用户别指望它一开始就「懂事」，记忆是跑出来的。前几个任务最好只读跑通、确认逻辑再放开，不然权限一宽误操作影响面不小。整体方向我看好，值守类重复活交给它挺合适。

---

• 

  莲花280

  —

  并行执行是真的爽，可以同时给好几个数字员工派活，不用像串行那样排队等。组建一支小团队的感觉很奇妙。

---

• 

  掠影_12

  —

  自定义财务那个岗目前还是半成品，它能帮我审合同找风险点，但没跟电子税务局打通，真正报税还得自己来。权限问题解决了，AI 财务能干的事才多。不过光审核合同这一项已经值回票价了。

---

• 

  LejlaNoel

  —

  记忆沉淀这个设计，越用越懂你项目，有点东西。

---

• 

  JordanGonzales_2022

  —

  和 QoderWork 的区别我算整明白了：Work 是白天帮你干杂活的效率工具，Wake 是下班后替你值守的守夜人，专门处理线上告警、代码审查这种活。要运维值守就选它。

---

• 

  李君超

  —

  交付不直观是真痛点，成果都在本地文件夹，页面上不展示，没指定工作空间根本找不到文件。

---

• 

  郭妍_1

  —

  内容运营那个数字员工有点东西，丢一篇公众号链接进去直接给我出了 8 张适合发小红书的信息流图，还顺手分析了推荐流公式。给它装了 20 多个 Skill 之后能力明显上来了，工作记录、记忆、新学的技能在页面都能追。

---

• 

  Mark_Cooper_Plus

  —

  接了钉钉之后 @它就能派活，响应比等同事回复快多了，复杂需求还是自己来。

---

• 

  Jonathan208

  —

  Mac -only 这点劝退，等 Windows 版。

---

• 

  Nicole.Morales_Pro7

  —

  说个真实场景：CI 挂了之后它自动拉日志、定位失败的 Job、还把报错和最近一次提交关联起来给根因分析，第二天早会前我就不用自己翻日志了。不过它给的报告里偶尔会漏掉一些不明显的目录，还是得人扫一眼。

---

• 

  DSmithK9

  —

  一开始只敢给只读权限，跑了一周发现逻辑挺稳，才慢慢放开写入。建议新手都这么来，别一上来就全开。

---

• 

  Stephanie_Sanchez520

  —

  我们团队给它配了个数字程序员，PR 合并自动出变更简报，省了不少同步的功夫。

---

• 

  Arthur.JimenezQ

  —

  半夜不用爬起来看日志了，真香。

No reviews yet