When software's producer shifts from humans to agents, how should the cloud itself be redesigned?
1. Vibe Coding platforms are reshaping how software is made
Over the past 18 months, a brand-new product form has moved from the fringe to the mainstream: the Vibe Coding platform.
Users no longer write code, configure environments, or even face an IDE — they just say one sentence in a dialog, like "make me a to-do list" or "make a website where friends can leave me messages", and a few minutes later they have an accessible, database-backed, login-capable online app.
This track has jumped from "toy" to "serious business" in the past six months, with a few landmark numbers:
- Lovable: ARR passed $400M in February 2026, with just 146 people and $100M in new monthly revenue; valuation climbed from $1.8B to $6.6B;
- Vercel v0: pushed Vercel's overall ARR to $340M with 84% YoY growth, closing a $300M Series F in March 2026 at a $9.3B valuation;
- Replit: grew revenue from $10M to $100M in 9 months on Replit Agent, closing a $400M Series D in March 2026 at a $9B valuation;
- Loopit: a domestic AI-native app incubated by Baichuan Intelligence's co-founding team, closing a $50M new round in April 2026.
If IDE coding agents like Cursor, Claude Code and CodeBuddy changed professional developers' workflows, Vibe Coding platforms are doing something more radical: lowering the barrier of "owning software" to what ordinary people can reach.
2. The engineering challenge today: LLM code-writing is no longer the bottleneck
A Vibe Coding platform looks like an AI product on the surface, but underneath it's a composite of agent runtime, cloud sandboxes, app backends, multi-tenant isolation and pay-as-you-go billing. For developers, the real challenge isn't getting the model to write code — it's organizing these engineering capabilities into a stable, secure, low-cost, reusable infrastructure foundation.
From an engineering view, it's a stack of several complex systems:
Agent Loop × multi-tenant cloud sandbox × complete app backend × strict permission and billing
Any team that wants to ship a Vibe Coding platform must solve four problems:
| Vibe Coding challenge | Infrastructure needed |
|---|---|
| How the agent runs and keeps working | Recoverable runtime, isolated sandbox, event logs, session state, context management |
| How the app becomes genuinely usable | Database, auth, storage, hosting, gateway |
| How multi-tenancy isolates | Tenant-level resource boundaries, permission model, pay-as-you-go |
| How to keep operating costs down | Serverless elasticity, scale-to-zero, storage tiering, lifecycle management |
Few teams can build these four from scratch — they correspond to underlying capabilities cloud vendors have polished for over a decade.
3. CloudBase's Vibe Coding platform solution
CloudBase is Tencent Cloud's AI-native full-stack app development platform, designed for AI-driven development flows and end applications, providing auth, database, cloud functions and containers, file storage and web app hosting as complete serverless resources.
3.1 Keep the agent running safely, stably and recoverably
CloudBase separates the Agent Loop, Sandbox, Session and controlled boundary: the Agent Loop orchestrates models and tools, the Sandbox executes untrusted code, Session records events and state, and sensitive operations go through the MCP Proxy, management-plane API and temporary credentials. This avoids mixing model inference, code execution and platform credentials in one runtime environment.
3.2 Make agent-generated apps genuinely usable
A Vibe Coding artifact shouldn't stop at a static page — it should naturally get database, auth, cloud storage, cloud functions, static hosting, container hosting and HTTPS access. CloudBase wraps these backend resources into tools the agent calls via MCP, letting the agent create and configure resources declaratively.
3.3 Make the agent faster, cheaper and safer
Treating the agent as a new kind of "developer", CloudBase redesigned several dimensions from an "agent's perspective":
- MCP tools replace the SSH + Bash operation chain: login config, table creation, permission rules, domain binding are all wrapped into structured MCP tools;
- Structured returns replace raw stdout/stderr: MCP tools return JSON with error codes and next-step suggestions, not system logs;
- BaaS lets the agent "call" rather than "build" the backend: Auth, cloud database, cloud storage, cloud functions out of the box;
- Platform-level security boundary: HTTPS API egress only, no SSH exposure, platform-default DDoS protection.
An equivalent controlled experiment on a Todo app, with the same agent and prompts, along "traditional VM deploy" vs "CloudBase" paths:
| Dimension | Traditional VM deploy | CloudBase + AI dev | Advantage |
|---|---|---|---|
| Completion time | 990s (16min) | 260s (4min) | 3.8× faster |
| Tool calls | 79 | 36 | 2.2× fewer |
| Agent internal loops | 189 | 89 | 2.1× fewer |
| Token usage | 2,788,291 | 1,323,431 | 2.1× cheaper |
| Public attack surface | SSH 22 + HTTP 80 exposed | HTTPS API only | SSH surface eliminated |
Core conclusion: at feature parity (anonymous sessions + Todo CRUD + data isolation), the CloudBase AI path is 3.8× faster than traditional VM deployment, cuts token consumption by 52%, and eliminates the security risk of SSH exposure.
The most direct source of these number differences is the agent's "attention allocation": on the VM path, the agent spent 76% of tool calls on ops like SSH, file transfer and process management; on the CloudBase path that dropped to 14%, with the remaining 78% returning to "understand the requirement, write the code" — what the agent is genuinely good at.
3.4 A unified multi-tenant model for isolation, cost and ease
CloudBase uses an "N+1 architecture": one platform environment hosts the Agent Loop, Sandbox and model inference; N user environments host each user's generated app. Each user environment has independent data, compute, storage and access boundaries, with pay-as-you-go and platform integration.
3.5 OpenVibeCoding: a fully open-source Vibe Coding platform built on CloudBase
To further lower the entry barrier, CloudBase implemented a fully open-source OpenVibeCoding as a starting point for teams to validate architecture and do secondary development.
Project: https://github.com/TencentCloudBase/OpenVibeCoding
- Out of the box: deeply integrated with CloudBase, AI can directly operate the database, cloud functions and storage;
- Brain/Hands separation: an advanced Harness architecture that raises task completion;
- Multi-tenancy: multi-user/task isolation, second-level environment creation, usable as SaaS directly;
- Production-grade: sandbox preview, HMR hot reload, integrated mini program toolchain.
4. Real customer cases
4.1 Tencent Tusi APP
Tusi APP (https://tusi.qq.com/) is the "app generation and idea co-creation platform" from the Tencent App Store team. Tusi uses CloudBase as its main backend:
- Cloud database: every user-generated app auto-creates an independent instance, isolated at the tenant level by nature;
- Auth: WeChat, phone, email and anonymous login built in;
- MCP / Skills: backend resources wrapped as standard MCP tools for declarative agent invocation.
4.2 GenieAI (CodeBuddy)
GenieAI (https://genie.codebuddy.ai/) is the online AI programming workbench from Tencent Cloud CodeBuddy, more pro-sumer: generated projects are complete full-stack apps with front end, back end and database.
- PostgreSQL: native transactions, foreign keys, complex SQL, vector search, instance-level tenant isolation;
- Auth (Google login): for global users, out of the box;
- One-click deploy and white-label domains: HTTPS link in seconds, custom domain binding supported.
5. Agents are the new developers — the cloud must be redesigned for them
In the future, it won't mainly be human engineers issuing commands, calling resources and deploying code to production — it'll be large numbers of concurrently running, tireless AI agents.
Accept that, and the conclusion is almost inevitable: a cloud designed for human developers is not a cloud designed for agents.
| What an agent-friendly cloud needs | CloudBase's approach |
|---|---|
| Backend callable by agents directly | All capabilities exposed as MCP tools, with matching Skills |
| Docs that fit the context window | LLM-optimized concise docs and llms.txt |
| Second-level cold start, use-and-leave | Full-stack serverless, scale-to-zero, millisecond startup |
| Secure sandbox for untrusted code | CloudBase Sandbox, isolated execution, snapshot recovery |
| Simple, unified model inference | Built-in AI model gateway, unified auth, billing, rate limiting |
Individually, none of these is revolutionary; together, they answer a bigger question: when software's producer shifts from humans to agents, how should the cloud itself be redesigned?
6. Closing
Recapping what this post wanted to say:
- Vibe Coding platforms are the main battleground of AI programming's second half — the question is no longer "let AI write code" but "let AI autonomously produce usable software";
- Supporting this new form needs a complete infrastructure: Agent Runtime, secure sandbox, app backend, multi-tenant isolation — none can be missing;
- This infrastructure has its own design principles: Brain/Hands separation, controlled trust boundaries, one tenant per environment, pay-as-you-go, MCP standardization;
- Tencent Cloud CloudBase provides the complete solution, with OpenVibeCoding as the open-source reference implementation.
