Vercel Toolbar now includes "Copy for Agents" functionality that captures complete visual context from comments, providing coding agents with the technical details they need to understand deployment feedback across your application.
When teams copy comments using this feature, agents receive structured context including page URL and viewport dimensions, selected text and node path information, React component tree details, and the original comment text. This helps agents understand exactly where issues occur in your deployed application and what changes are needed.
This structured format eliminates the need to manually explain deployment context to agents, enabling direct copying from the toolbar with complete technical details for component location and implementation.
The feature is available to all Vercel users immediately.
Workflow 4.1 Beta changes how workflows track state internally. Instead of updating records in place, every state change is now stored as an event, and current state is reconstructed by replaying the log. This release also adds support for provider-executed tools and higher throughput.
Event sourcing is a persistence pattern where state changes are stored as a sequence of events rather than by updating records in place. Instead of storing "this run is completed," the system stores "run_created, then run_started, then run_completed" and reconstructs the current state by replaying those events.
In Workflow 4.1, runs, steps, and hooks are no longer mutable database records. They're materializations of an append-only event log. Each event captures a timestamp and context, and the runtime derives current state by processing events in order.
This architecture makes workflows more reliable in three ways:
Self-healing: If a queue message is lost or a race condition occurs, replaying the workflow route detects missing state and re-enqueues the necessary messages. Old runs required manual intervention to recover from queue downtime; new runs recover automatically.
Complete audit trail: The event log lets you replay the exact sequence that led to any state, which makes debugging distributed workflows much easier.
Consistency: Events are append-only, so partial failures during a write can't leave entities in an inconsistent state. The event log is the single source of truth.
For a deeper look at the event model, including state machine diagrams for run, step, and hook lifecycles, see the Event Sourcing documentation.
Improved throughput: The workflow queue system now processes many thousands of steps per second. When dependencies allow, multiple steps execute in parallel.
Provider-executed tools: @workflow/ai now supports provider-executed tools like Google Search and WebSearch, which run on the model provider's infrastructure rather than in your workflow.
NestJS support: The new @workflow/nest package adds build support for NestJS applications, handling dependency injection patterns so workflows integrate with existing NestJS services.
Top-level using declarations: The SWC plugin now supports the TC39 Explicit Resource Management proposal inside step and workflow functions, enabling automatic resource cleanup.
Custom class serialization: Client mode now supports custom class serialization, with a classes object in manifest.json that declares serializable types.
Fixed double-serialization of tool output in @workflow/ai
Vercel now supports Koa applications, an expressive HTTP middleware framework to make web applications and APIs more enjoyable to write, with zero-configuration.
server.ts
import Koa from'koa'
import{ Router }from'@koa/router'
const app =newKoa()
const router =newRouter()
router.get('/',(ctx)=>{
ctx.body ={ message:'Hello from Koa!'}
})
app.use(router.routes())
app.use(router.allowedMethods())
app.listen(3000)
A "Hello World" Koa app on Vercel
Backends on Vercel use Fluid compute with Active CPU pricing by default. This means your Koa app will automatically scale up and down based on traffic, and you only pay for what you use.
Builds and Functions now support Python 3.13 and Python 3.14 alongside the previously supported Python 3.12. Projects without a specified Python version continue using Python 3.12 by default.
The default will switch to Python 3.14 in the coming months. To continue using Python 3.12, specify an upper bound in your project manifest (pyproject.toml or Pipfile) as shown in the examples below.
Vercel Sandboxes are now generally available, providing an ephemeral compute primitive for safely executing untrusted code.
It lets teams run AI agent-generated outputs, unverified user uploads, and third-party code without exposing production systems.
Each sandbox runs inside Firecracker microVMs, isolated from your infrastructure, so code running in a sandbox is blocked from accessing environment variables, database connections, and cloud resources.
Sandboxes are in production use by teams including v0, Blackbox AI and RooCode.
To bootstrap a simple Node.js application that creates a Vercel sandbox, use the code below:
import{Sandbox}from'@vercel/sandbox';
const sandbox =awaitSandbox.create();
await sandbox.runCommand({
cmd:'node',
args:["-e",'console.log("Hello from Vercel Sandbox!")'],
stdout: process.stdout,
});
await sandbox.stop();
Or get started with the CLI by opening an interactive shell:
The Vercel Agents Marketplace now includes cubic, an AI code reviewer that that deploys thousands of AI agents to find and fix bugs in your PRs and codebase.
Most code review tools only see what changed. cubic sees how those changes connect to everything else. It learns from your team’s past reviews and gets better over time.
Key capabilities include:
Catching bugs, regressions, and security vulnerabilities in PRs and existing codebases; continuously running 1000s of agents
Identifies senior engineers on your team and learns from their comment history
Applying fixes automatically through background agents
With cubic handling the first pass, teams spend less time on manual review and more time merging changes. Custom coding standards get enforced across repositories, helping keep code consistent as teams scale.
The integration connects natively with Vercel, so adding AI-driven customer support takes minutes. With AssistLoop, teams can:
Install AssistLoop with minimal setup using an Agent ID
Add AI-powered support directly to Next.js apps
Train agents on internal docs, FAQs, or knowledge bases
Customize the assistant to match your brand
Review conversations and hand off to human support when needed
This integration fits naturally into existing Vercel workflows, with unified billing, automatic environment variables, and no manual configuration. Teams can ship AI-powered support faster without managing separate dashboards or complex setup.
AssistLoop automatically injects NEXT_PUBLIC_ASSISTLOOP_AGENT_ID into your project environment. Add the widget script to your site:
+2
