For many programmers, this scenario is familiar: a red error message pops up in the terminal, the log is copied, pasted into the AI assistant, analysis is awaited, and then the solution is pasted back into the terminal to run… If the problem isn't solved the first time, the next cycle begins. This is the daily reality for many Vibe Coding users.
Over the past few years, AI programming tools have evolved rapidly, and development efficiency has continuously improved. However, at the same time, a new contradiction has gradually emerged:Developers are increasingly becoming its "nannies," pinned to their computers and manually moving context throughout the process.
The reason for this is that these AI programming tools have always remained outside the R&D process, disconnected from the development team. They can write code, but they cannot act like a real team member, taking on a task and completing the entire process from investigation to delivery.
Recently,Tencent Cloud officially released CodeBuddy NPC, a cloud-based intelligent agent.This aims to provide a solution to the aforementioned problems. Developers simply need to @ an NPC on the cloud development platform, just like @ a team member, to assign tasks. The NPC can then autonomously complete problem investigation, solution formulation, code development, and PR submission, and continuously fix issues based on CI results until an acceptable result is delivered.
CodeBuddy NPC is the industry's first product to propose and implement the AI Native Git paradigm. Behind it lies Tencent Cloud's evolving technological philosophy, from CodeBuddy and WorkBuddy to CodeBuddy NPC.Integrate AI into real workflows.
Prior to the release, Zhidongxi had in-depth discussions with Tencent Cloud CodeBuddy, Huang Youkun, CNB product manager, Wang Chunyu, CNB operations manager, and Xu Yi, CNB senior engineer, in an attempt to understand the technical logic and industry assessment behind the product.
01.
From playing support to taking on key roles
What can CodeBuddy NPCs do?
According to Wang Chunyu, the evolution of AI’s role in the R&D field can be divided into three stages: the first stage is AI-assisted R&D, where humans are still the dominant force and AI only performs single-point tasks such as code completion, generation, and question-and-answer explanation.
As the capability gap between domestic and overseas large-scale models continues to narrow by 2025, the industry will enter the second stage of AI-driven development, where AI will begin to take over local tasks, such as fixing bugs, conducting tests and verifications, and integrating into pipelines to fix defects.
The third stage is AI engineering collaboration, where AI is no longer just an efficiency-enhancing tool, but has its own functional positioning, knows how to participate in the R&D process, and can complete tasks in a closed loop and deliver business value independently.
Wang Chunyu judged:The industry is currently in a transitional period from the second to the third stage."CodeBuddy NPC is a product created for the third phase."
In terms of capabilities, NPCs can autonomously complete tasks from...Task acceptance, planning, execution, verification to iterative optimizationThe entire process: After receiving a task from an Issue, it will automatically read the requirements and code, provide a solution, complete the development, and submit a PR; when encountering code merge conflicts, it will handle them automatically and resubmit; after CI fails, it will automatically locate the problem, fix the code, and re-execute until it passes the quality gate.
NPC targets high-frequency scenarios in team R&D, such as defect fixing and new feature development, and streamlines the entire delivery chain from requirement proposal and problem analysis to code submission and merging.
Compared to the traditional Vibe Coding model, which involves continuous developer involvement and repeated debugging, NPC further reduces the cost of human intervention: defect fixing has shifted from a process requiring multiple rounds of collaboration between humans and AI, taking several hours, to a process that developers can directly address.Complete result acceptance within minutes.The development of new features, which used to take several days, has also been...Shortened to "hourly"Developers can mainly focus on the requirements confirmation and code acceptance stages.
It's worth mentioning NPC's "on-demand summoning" mechanism. In the interview, Huang Youkun specifically clarified its difference from common Agent products: many Agents need to be created, authorized in advance, and run continuously, which is equivalent to "raising" one before it can be used.
Before being @mentioned, an NPC is merely a YAML definition in the repository and doesn't actually exist. Only when it's invoked does the platform dynamically create a working environment, loading the YAML-defined roles, prompts, and execution strategies into a Docker sandbox for execution, and then destroying it once the task is complete. This allows developers to @mention multiple NPCs simultaneously to work in parallel, or assign different roles to NPCs.
▲YAML file defining NPCs and different NPC roles
The CodeBuddy NPC team showcased a compelling case study: an NPC team completed the entire development process of a mini-game called "Signal," from requirements decomposition to final verification, without any human intervention. The project manager NPC was responsible for task breakdown and allocation; scriptwriting, art, music, and voice acting NPCs each created their content; the asset integration NPC compiled the code and materials; and finally, code review and testing NPCs completed the verification. The output of one NPC directly became the input of the next.
Regarding the relationship between NPCs and local IDE tools like CodeBuddy, the CodeBuddy NPC team's stance is clear: coexistence, not replacement. Huang Youkun used an analogy: AI in an IDE is more like making a phone call—synchronous, requiring multiple rounds of interaction and clarification of requirements; while NPCs are asynchronous—once the task is defined, it goes to work after being mentioned."You can even turn off your computer and go home, and it will still be running in the cloud."
02.
CNB platform has accumulated 8 years of experience.
AI has become a "native" part of the R&D process.
The core of CodeBuddy NPC is the AI Native Git paradigm proposed by Tencent Cloud.
In the Vibe Coding model, context is handled by people; however, in a real R&D team, this context already exists within the process: Issue is the requirement memory, recording tasks, background, and goals; code repository is the engineering memory, carrying project structure, dependencies, and evolution history; PR is the change memory, accumulating solutions, code, and review comments; CI is the quality memory, recording build results, reasons for failure, and fixes.
The idea behind AI-native GitThis means enabling AI to use the Git repository as its native memory, directly integrating it into the team's existing Git, Docker, CI/CD, and quality access control systems.This allows for the reuse of existing pipelines to form an automated closed loop of "submission-verification-repair-reverification". Enterprises do not need to modify their toolchains, and AI can autonomously perceive the context.
The foundation supporting this paradigm isTencent Cloud's Cloud Native Building Platform (CNB)The CNB platform has been iterating internally at Tencent since 2018, accumulating experience for over eight years, and was launched for public beta in 2024.
In the era of large models, the addition of model files has caused the size of repositories on CNB to expand dramatically, with many repositories exceeding 100GB and even reaching TB levels. How to achieve code cloning in seconds under ultra-large repositories and quickly build Docker image environments within container sandboxes are capabilities that CNB has focused on refining in recent years, and now they have become prerequisites for NPCs to quickly create verification environments and start development tasks.
In addition, CNB provides token measurement and billing capabilities, making it clear how many tokens each task consumes, which NPC uses them, and how the budget is managed. Wang Chunyu revealed that...CNB currently boasts over 100,000 monthly active developers and a daily retention rate as high as 80%, and is already being used by numerous B-end enterprise clients.
For enterprise managers, another value lies in the record-keeping. Huang Youkun frankly stated that what developers say to AI in the local IDE is their personal asset, but the cost is borne by the enterprise. Once a problem occurs after deployment, it is difficult to trace back "why the AI wrote the code the way it did".
On CNB, NPCsThe entire reasoning process, execution rounds, calling model, and token consumption are auditable and traceable.The AI production process is thus incorporated into the enterprise's assets, something that local IDEs struggle to achieve in terms of architecture.
Cost control presents another engineering hurdle. Each agent task execution involves multiple rounds of model calls. The initial system prompts and tool descriptions are reused in subsequent calls; the smaller the initial token, the lower the overall cost. This is achieved through continuous optimization of prompts, tool calls, CLI output, and cache hit rate.The initial token consumption for official NPCs has decreased from over 20,000 in the early stages to approximately 2,000, a reduction of over 90%.
When discussing how CodeBuddy NPC differentiates itself from other AI programming products, Huang Youkun offered a formula: organizational effectiveness equals effectiveness divided by friction.AI acts as an amplifier, but friction in the process can diminish the amplifying effect.For example, if a CI error occurs and someone needs to handle it, the loop can idle for 20 minutes over the time it takes to eat a meal; if the local AI needs authorization before executing commands, the task will freeze if the user is not at the computer. In his view, the elimination of friction through platform-based infrastructure is the real differentiator for products like CodeBuddy NPC.
03.
After AI is integrated into workflow
The role of "human" has also changed.
Analyzing Tencent Cloud's product line reveals a clear trajectory: from CodeBuddy to WorkBuddy, and now to CodeBuddy NPC, Tencent is not simply distributing AI tools to everyone, but rather attempting to deliver a collaborative approach restructured around AI. While the scenarios for these products differ slightly,However, it maintained the basic idea of "integrating AI into real workflows".
Initially, CodeBuddy started with code assistant plugins, and then expanded to various forms such as IDEs and CLIs, constantly getting closer to the daily work environment of developers, and achieving large-scale application within Tencent.
▲CodeBuddy in IDE form
In March 2026, WorkBuddy, created by the CodeBuddy team, further expanded this concept from R&D scenarios to a wider range of office collaboration scenarios. Through capabilities such as workspaces, expert teams, skill systems, and tool connectivity, it enables AI to participate in specific tasks such as document processing and content creation.Explore the shift from "AI-assisted office work" to "AI performing tasks".
The newly released CodeBuddy NPC further delves into the software development chain, focusing on upgrades around Harness (the cloud-based intelligent agent runtime governance layer).Transform "AI teammates" into direct participants in the R&D process.
Within Tencent, this philosophy has evolved into a kind ofThe collaborative closed loop of the "AI native team"PM initiates tasks through the Requirements Input Skill, with AI participating in requirement clarification and task breakdown; Technical Leaders and AI co-build technical solutions and generate R&D tasks; R&D and AI collaborate to complete development verification; Delivery Analysis Skill collects data from the entire process and feeds back into the next round of process optimization.
This signifies a fundamental shift in the developer's role:From code producers to supervisors and rule makers who set goals, make decisions, and verify results.In the interview, Huang Youkun described a specific scenario: With NPCs, developers can assign all the day's tasks to the NPCs as soon as they arrive at work, and then collect all the "work" after a while to decide whether to modify, rework, or merge it. "The proportion of human involvement has decreased, but people must still oversee it, because this code will ultimately affect the online service."
Of course, bottlenecks also exist, and Tencent does not shy away from acknowledging this.One issue is that AI generates public relations documents (PRs) too quickly, and human review and acceptance are becoming a new efficiency bottleneck.Huang Youkun said, "There are probably a lot of pull requests piling up in the repository waiting for developers to review. This problem has not been completely resolved yet."
The current feasible solution is to improve the level of automated verification, allowing AI to self-repair based on test and scan results, moving towards a closed loop of 24/7 unattended operation. In the future, even when people are sleeping, NPCs can collaborate with each other to push the project forward.
Secondly, the effectiveness of these tools depends heavily on the level of automation infrastructure provided by the enterprise itself.Teams with high engineering skills and a high degree of automation will see their benefits amplified rapidly; teams with low levels of automation will need to improve their infrastructure first before they can fully unleash the power of AI.
04.
Conclusion: R&D process reengineering is a new battleground for AI programming.
The release of CodeBuddy NPC, viewed within a broader industry context, sends a clear signal: the focus of competition in AI programming is shifting from "how well and how fast the model can write code" to "whether AI can be embedded into real-world R&D processes." The former is a tool-based logic for improving individual efficiency, while the latter is an infrastructure-based logic for organizational collaboration.
Following this line of thought, the actions of the leading players are converging. Cursor launched Origin, a code hosting platform for agents, extending its reach from editors to code repositories; OpenAI is reportedly developing an internal alternative to GitHub; and Anthropic introduced the Claude Tag, allowing developers to delegate tasks to Claude on Slack as if they were calling out to a colleague. While their paths differ, they all point to the same goal: transforming AI from a tool for writing code into a collaborative teammate.
This transformation won't happen overnight. AI-generated PRs are already faster than human reviewers, and the level of a company's automation infrastructure directly determines its revenue ceiling. Review bottlenecks and infrastructure barriers still need to be overcome. But the direction is clear: restructuring R&D processes is becoming the next major battleground for AI programming.