Vercel AI SDK 7 transforms from a simple model wrapper into a comprehensive agent platform. Learn how WorkflowAgent, SandboxSession, and approvals enable production-ready TypeScript agents.
Imagine a production server crashing at two in the morning because a serverless function timed out while streaming a complex, multi-step language model response. Or picture a user closing their browser tab right in the middle of a database-writing agent execution, leaving your application state completely corrupted. These are the exact operational headaches that client teams bring to us when they attempt to scale past a simple chat interface and build real-world agentic systems.
For the last couple of years, building artificial intelligence features meant throwing together a frontend chat window and calling an API. But as we move deeper into 2026, the industry is transitioning from toy chatbots to autonomous agentic workflows that actually execute tasks. To address these challenges, Vercel officially released Vercel AI SDK 7 on June 25, 2026. This major update marks a significant evolution, transforming the library from a simple model wrapper into a comprehensive runtime for building production agents in TypeScript.
Whether you are a technical founder, an engineering lead, or a product manager, this guide will walk you through what has changed. We will explore the new primitives, analyze the architectural shifts, and help you decide how to integrate these upgrades into your product roadmap.
Vercel AI SDK 7 is an open-source TypeScript framework that transitions from simple model wrappers into a comprehensive agent platform. It introduces first-class abstractions for durable execution, human approvals, sandboxed runtimes, and multi-agent coordination, enabling developers to build resilient, production-ready AI agents that survive system restarts and redeploys.
By shifting the focus from simple token streaming to robust execution state management, this release standardizes the backend infrastructure needed to run agents reliably. Instead of writing thousands of lines of custom state-persistence and verification code, engineering teams can now use native SDK primitives to handle multi-step workflows, secure code sandboxes, and manual approval gates.
Early versions of the Vercel AI SDK focused almost entirely on frontend user experience. Primitives like the useChat hook made it incredibly easy to stream text responses from providers like OpenAI or Anthropic directly into a React frontend. This was highly effective for building simple chatbots, but it fell short when teams attempted to build actual agents that need to make decisions, execute tools, and run over multiple turns.
As web applications became more sophisticated, developers had to write complex orchestration loops. They had to manually manage conversation history, handle API retry logic, and parse structured tool outputs. While Python-based frameworks like LangChain, CrewAI, or AutoGen dominated this space, they created a massive disconnect for full-stack teams. Writing agents in Python meant maintaining a separate microservice, dealing with cross-language serialization, and losing the tight integration with their Next.js or React frontends.
Vercel AI SDK 7 is a direct bid to make TypeScript the premier language for production-ready agent development. It is not just about calling a model. It is about providing a complete runtime environment that manages the life cycle of an agent from start to finish. This means handling execution state, tool calling, environmental sandboxing, and observability in a unified, type-safe manner.
The Vercel AI SDK has climbed to over 16 million weekly npm downloads, securing its position as the standard library for TypeScript-based AI development.
This rapid adoption reflects a broader industry trend. Developers are tired of fighting fragile, multi-layered abstractions. They want simple, idiomatic tools that integrate naturally with modern web frameworks. By standardizing these patterns, this release helps teams move past the experimental phase and build systems that can be trusted in production environments.
One of the most significant engineering challenges in agent development is handling long-running executions. If an agent has to execute a multi-step plan, it might take several minutes to complete. It might need to search the web, write a file, run a test, and update a database. If the server hosting the agent crashes, or if you deploy a bug fix to production while the agent is on step three, the entire execution is typically lost. The user gets a generic error and has to restart the entire process, which is frustrating and wastes expensive tokens.
The solution in Vercel AI SDK 7 is the new @ai-sdk/workflow package and the WorkflowAgent class. This introduces durable, resumable execution to the TypeScript agent ecosystem. Instead of running the entire agent loop in volatile memory, WorkflowAgent breaks the execution down into discrete, stateful steps.
Between each step, the agent automatically serializes its complete state, including the conversation history, the tool execution results, and the model's internal reasoning traces. This state is persisted to a durable storage layer. If the process is interrupted by a server restart, a network drop, or a new deployment, the agent can be rehydrated and resume execution from the exact step where it stopped.
This level of durability is the foundation of production-grade agentic systems. In our strategy guide on AI coding agents in production, we detailed how execution durability is the single most critical factor for enterprise adoption. Without it, your agents are simply too fragile to be trusted with mission-critical workflows. By integrating durability directly into the SDK, Vercel has removed one of the largest architectural hurdles facing development teams today.
As agents become more autonomous, security and control become paramount. If you build an AI assistant for an enterprise platform, you might give it tools to modify database records, send emails, or charge credit cards. If the model hallucinates or is manipulated via a prompt injection attack, letting it run completely unattended is an absolute recipe for disaster.
To prevent this, you need a human-in-the-loop system. The agent must pause its execution and ask a human for permission before running a high-risk tool. Before Vercel AI SDK 7, building this was a monumental headache. You had to save the agent's state to a database, send an email or Slack notification, expose an API endpoint for the approval, and then somehow rehydrate the agent and resume the loop when the approval came in. It required hundreds of lines of custom, fragile boilerplate code.
With the AI SDK 7 release, tool approvals are now a first-class primitive. When defining a tool, you can simply flag it as requiring manual approval. When the agent attempts to call that tool, the SDK automatically pauses the execution loop, serializes the current context, and triggers an approval event.
The frontend can listen for these events and render a clean, interactive approval interface. The user can review the proposed parameters, modify them if necessary, and either click approve or reject. Once the decision is made, the frontend sends the payload back to the server, and the SDK resumes the agent loop.
This structural safety is essential for modern security compliance. We explored this threat landscape deeply in our analysis of how the agentjacking exploit redefines security. Unmonitored, auto-executing tools are a massive security vulnerability. By integrating native, durable approvals, SDK 7 makes it significantly easier to secure your applications against unauthorized actions.
If you are building an AI developer assistant, a data analysis tool, or an automation bot, your agent will often need to write and run code. For example, it might write a Python script to clean a CSV file and then execute it to get the results. Running that code directly on your production servers is an absolute security nightmare. A user could trick the model into executing a command that deletes files, accesses sensitive environment variables, or contacts external servers.
To prevent this, you must run the code in an isolated sandbox. But managing these sandboxes is incredibly complex. The environment needs to be fast to spin up, secure, and consistent across local development, CI/CD pipelines, and production environments.
Vercel AI SDK 7 introduces the SandboxSession abstraction to solve this. It provides a standardized, portable interface for executing terminal commands and running code within secure, isolated sandboxes. The beauty of this abstraction is that it completely decouples your tool definitions from the underlying sandbox provider. You can write your tools using a generic SandboxSession interface, test them locally using a local Docker setup, and then swap the provider to Vercel Sandbox or E2B in production by changing a single line of configuration.
This portable isolation is exactly how we approach infrastructure safety for our clients. When we consult on product architecture, we ensure that dynamic execution environments are strictly isolated from core business data. Having a standardized sandbox layer makes it much easier to build AI products that survive sudden model shutdowns and other runtime anomalies without exposing your infrastructure to security exploits.
Model Context Protocol (MCP) is an open standard developed by Anthropic that allows agents to connect to external data sources, APIs, and developer tools. It has quickly become the industry standard for tool integration. However, Vercel identified a key limitation in the original MCP design. It was built purely for text-based interactions. The model requests data, the tool returns text, and the model processes it. But in a modern web application, users do not just want to read text. They want to interact with data. They want charts, forms, and interactive widgets.
some tools require a specialized user interface just to configure them or display their outputs, and these interfaces should be isolated from the language model itself to prevent prompt injection or UI spoofing.
AI SDK 7 addresses this by introducing MCP Apps. This extension allows MCP servers to not only expose tools to the language model, but also to expose specialized user interfaces to the application. These app UIs are rendered inside secure, sandboxed iframes on the frontend. A structured JSON-RPC bridge handles the communication between the agent, the underlying tools, and the interactive display.
For example, if an agent is helping a user manage their calendar, it can use an MCP tool to fetch events, and then render an interactive calendar widget. The user can drag and drop events directly on the screen, and those changes are communicated back to the agent.
This represents a massive leap forward in UI/UX design. In our analysis of how the latest AI and mobile upgrades are shifting product roadmaps, we emphasized that the future of interfaces is not static layouts. It is dynamic, context-aware environments assembled on the fly by agents, and MCP Apps are a major step toward making that a reality.
On June 29, 2026, Vercel announced native support for audio and voice on its AI Gateway, fully integrated into AI SDK 7. Most existing voice interfaces are incredibly clunky. They feel like a series of walkie-talkie exchanges rather than a natural conversation. This is because they chain multiple separate models together: a speech-to-text model transcribes the user's voice, a text-based language model generates a response, and a text-to-speech model synthesizes the audio.
This multi-step pipeline introduces massive latency. The user speaks, and then has to wait several seconds for the system to process, think, and generate the audio. It also loses all the nuances of human speech, such as tone, emotion, and the ability to interrupt.
AI SDK 7 supports native, end-to-end realtime voice models (such as those from OpenAI and xAI) routed directly through the Vercel AI Gateway. These models process raw audio inputs and generate raw audio outputs directly, bypassing the clunky three-step pipeline entirely. This reduces latency to milliseconds, making the conversation feel truly natural. because it is a continuous stream, the user can interrupt the AI mid-sentence. The model immediately detects the user's voice, stops speaking, and listens, exactly like a human would.
To make this easy to build, the SDK introduces the useRealtime React hook. This hook handles all the complex browser-side logic: capturing microphone input, managing the WebSocket connection, and playing back the streaming audio. On the server side, you simply mint a short-lived session token via the AI Gateway, ensuring your master API keys never touch the client's browser.
This opens up incredible possibilities for customer support, language learning, and hands-free applications. As we discussed in our article on how AI developer agents shift your MVP scope, adding high-quality, zero-latency voice to your product is no longer a massive engineering hurdle. It is a feature you can implement in a matter of days.
When an agent runs across multiple steps, calling several tools and models, debugging becomes incredibly difficult. If the output is wrong, finding where it went off the rails is like finding a needle in a haystack. Was it a bad prompt, an incorrect tool output, or a model hallucination?
To solve this, AI SDK 7 integrates deep, structured telemetry built on top of the native Node.js tracing channels (node:diagnostics_channel). The SDK automatically emits structured execution events on the ai:telemetry channel for every major action, including text generation, model calls, tool executions, and embedding generations.
This means you do not have to write custom logging code for every single tool and agent step. Instead, any modern observability provider can subscribe to this channel once and automatically construct detailed, end-to-end execution traces. You can see exactly how many tokens were consumed at each step, what the prompt looked like, what the tool returned, and how long each operation took. This is invaluable for both debugging and performance optimization.
For example, when building complex search and retrieval systems, tracking latency is critical. Our guide on why you should choose pgvector over dedicated vector databases highlights how keeping your data layer simple and integrated makes it much easier to track and optimize your overall application performance. By pairing a unified data layer with the standardized telemetry of AI SDK 7, you gain complete visibility into your system's execution path.
While Vercel is positioning AI SDK 7 as a complete agent platform, it is important to compare it against existing players like LangChain, Mastra, and CrewAI. LangChain is a massive, multi-language ecosystem with hundreds of integrations. However, many developers find it overly abstract, heavily nested, and difficult to debug in TypeScript. It often feels like you are fighting the framework's abstractions rather than building your product.
AI SDK 7 takes a highly idiomatic, TypeScript-first approach. It integrates directly with React hooks, uses standard Zod schemas for structured outputs, and avoids deep inheritance patterns. It matches the mental model of modern web development and does not force you into complex, multi-layered class architectures.
Let us compare these two frameworks side by side:
| Feature / Metric | Vercel AI SDK 7 | LangChain (JS/TS) |
|---|---|---|
| Primary Language | TypeScript (Native) | Python & TypeScript |
| State Management | Durable WorkflowAgent |
StateGraph / LangGraph |
| UI Integration | First-class React Hooks | Manual stream parsing |
| Code Execution | Standard SandboxSession |
Custom tool execution |
| Telemetry | Native node:diagnostics_channel |
LangSmith (Dedicated) |
| Learning Curve | Gentle, idiomatic | Steep, heavy abstractions |
For teams building modern web applications, the AI SDK is almost always the more natural choice. It aligns perfectly with the mental model of modern frontend and full-stack development, allowing you to build highly interactive agentic features without the overhead of a massive, multi-language framework.
For teams running AI SDK v6 or older, upgrading is highly recommended to take advantage of these production-grade agent tools. To make this transition as smooth as possible, Vercel has provided automated migration tools. You can run the codemod command npx @ai-sdk/codemod v7 to automatically update your codebase. This codemod will analyze your files and rewrite deprecated model calls, structured output configurations, and tool definitions to conform to the new v7 API.
For more complex codebases, Vercel has also introduced a migration skill that can be added to your agent workspace using npx skills add vercel/ai --skill migrate-ai-sdk-v6-to-v7. This tool uses an AI-driven assistant to review your repository, identify potential breaking changes, and safely apply the necessary refactors.
When we handle these migrations for our clients as part of our custom software development services, we always recommend a staged approach. Start by creating an isolated feature branch and running the automated codemods. Then, set up a comprehensive test suite to verify that your custom tools and agent loops behave as expected under the new serialization and context-sharing layers before deploying to production.
While Vercel AI SDK 7 is free and open-source, the infrastructure costs of running complex agents can be substantial. Vercel's hosting platform was originally designed for static websites and short-lived serverless functions. They charge based on execution duration, measured in milliseconds.
But an agent is not a short-lived function. A durable WorkflowAgent that runs a multi-step loop, executes code sandboxes, and pauses for human approvals can remain active for several minutes or even hours. Even though the SDK serializes state and pauses execution, the active compute time spent managing the agent loop, streaming responses, and handling WebSocket connections for realtime voice can add up very quickly. If you host these workloads entirely on Vercel's serverless infrastructure, you may find your monthly bill scaling far faster than your user base.
This is why measuring the real ROI of AI is a critical step before launching any agentic feature. For early-stage startups and MVPs, the convenience and speed of deploying on Vercel are incredibly valuable. But as you scale, you need to consider the long-term cloud economics.
At Algoramming, we often help our clients build hybrid architectures. We keep the highly interactive frontend UI on Vercel to maintain a great user experience, but we deploy the heavy agent execution loops and background jobs on raw container compute, such as AWS ECS or EKS running on cost-effective Graviton instances. This approach gives you the best of both worlds: rapid frontend deployment and highly optimized, predictable infrastructure costs at scale.
For early-stage startups and MVPs, the speed of deploying on Vercel is unmatched. But as your daily active users grow, that convenience can turn into a substantial operational tax. A hybrid architecture, where the frontend UI remains on Vercel but the heavy agent execution loop is hosted on dedicated cloud compute, often yields the best balance of developer velocity and cost efficiency.
By decoupling the execution layer from the presentation layer, you can scale your computing resources independently. This is particularly important for CPU-intensive agent tasks, such as local data processing or running code sandboxes. It also allows you to implement custom pooling and caching strategies, reducing your overall token consumption and API costs.
Key takeaways
- Durable Execution: The new
WorkflowAgentensures agent executions can survive server restarts, redeploys, and network drops, making TypeScript agents far more reliable.- Human-in-the-Loop native support: Built-in tool approvals allow developers to easily implement safety gates for high-risk actions, pausing and resuming executions seamlessly.
- Isolated Code Sandboxes: The
SandboxSessionabstraction provides a standardized, portable interface for running generated code securely, decoupling tools from sandbox providers.- Realtime Voice on AI Gateway: Native support for low-latency, two-way audio conversations allows users to naturally interrupt the AI, transforming customer support and voice interfaces.
- Cloud Economics: While the SDK is free, hosting long-running agents on serverless infrastructure can cause bills to spike, making hybrid architectures on dedicated compute highly attractive.
Vercel AI SDK 7 transitions the framework from simple model wrappers and chat primitives into a complete, production-grade agent platform. It introduces native durability with WorkflowAgent, first-class tool approval gates, isolated code execution via SandboxSession, and deep, structured telemetry out of the box.
Yes, Vercel AI SDK 7 provides native support for multi-agent coordination. It allows you to build complex workflows where specialized agents can delegate tasks to one another, share state via typed runtime contexts, and coordinate execution steps under human oversight.
WorkflowAgent automatically serializes the complete agent state, including conversation history, tool results, and reasoning traces, between each execution step. This serialized payload is saved to a durable storage layer, allowing the agent to resume seamlessly after a process crash or redeploy.
Yes, the core Vercel AI SDK 7 library is entirely free, open-source, and provider-agnostic. While it integrates naturally with Vercel's platform, you can deploy your TypeScript agent code to any Node.js environment, container service, or public cloud platform without any lock-in.
When an agent attempts to call a tool flagged as requiring approval, the SDK automatically pauses execution and serializes the state. The frontend renders an approval interface, and once the user approves or rejects the action, the payload is sent back to resume the agent loop.
The SandboxSession is a portable abstraction that allows agents to execute terminal commands and run generated code within secure, isolated environments. It decouples the tool logic from the sandbox provider, allowing you to use local Docker setups or cloud sandboxes interchangeably.
Yes, Vercel AI SDK 7 introduces native support for realtime, end-to-end voice models routed through the Vercel AI Gateway. This allows developers to build zero-latency voice assistants where users can speak, listen, and naturally interrupt the AI mid-sentence.
Unlike LangChain's multi-language, heavily nested class hierarchies, Vercel AI SDK 7 is designed specifically for modern TypeScript and React ecosystems. It offers a much gentler learning curve, utilizes standard Zod schemas, and provides native, first-class React hooks for seamless UI integration.
Yes, Vercel provides an automated codemod utility. Running the command npx @ai-sdk/codemod v7 will automatically scan your codebase and rewrite deprecated model calls, structured output configurations, and tool definitions to conform to the new v7 API standards.
Building AI agents that work flawlessly in production requires a deep understanding of both software engineering fundamentals and modern cloud architecture. While Vercel AI SDK 7 provides the software primitives to build durable, secure, and observable agents, deploying them at scale requires careful planning around data storage, security boundaries, and infrastructure costs.
Choosing the right abstractions early in your product development life cycle will save your team months of refactoring and thousands of dollars in unnecessary cloud bills. Whether you are migrating an existing application or starting a greenfield project, focusing on execution durability, secure sandboxing, and cost-effective hosting is the key to building AI products that stand the test of time.
If you are planning a project like this and want to ensure your architecture is built for scale, we are happy to talk it through. Reach out to our team for a comprehensive tech partnership and consultation to evaluate your product roadmap and design an infrastructure that balances rapid feature delivery with long-term operational efficiency.
01 · RelatedDiscover how the Vercel Workflow SDK and the use workflow directive bring durable background jobs to Next.js without the complexity of traditional queues.
Read post
02 · RelatedCollaborative agentic workflows are here. Learn how to implement Vercel for Slack and the Cline AI SDK adapter to build secure, transparent, multi-agent systems.
Read post
03 · RelatedDiscover the critical lessons from Vercel's $1 Million Sandbox Challenge on HackerOne. Learn how to secure autonomous AI agents using microVMs, egress firewalls, and the Run SDK.
Read postWe will reply in plain English within one business day, NDA on request. Discovery call is free.
We design and engineer software, mobile, and web products end-to-end. Send the brief, we will reply within one business day.
Start a projectWe send a short email whenever we publish a new field note or ship a studio update. No fixed schedule, no filler.
Unsubscribe in one click. We never share your address.
