Get to know Cadenya
We’re developers who love to build. We set out to create a yes-code platform that makes building agents feel like the best parts of building software.
Vercel Workflow runs an AI agent as durable code: you mark the agent function with "use workflow", each model or tool call with "use step", and the Workflow SDK (first released as the Workflow DevKit) records every step so the run resumes after a crash or a deploy. A hook pauses the run for human in the loop, Vercel Sandbox runs the code the agent writes in a Firecracker microVM, and Cadenya is a hosted agent loop you can use instead of writing the loop as a workflow.
So pick Vercel Workflow when your app already deploys on Vercel, your agent is TypeScript, and you want the loop in your own code with WorkflowAgent from the AI SDK. Pick the Postgres World when you want the same code on your own servers. And if you’d rather not write, version and host the loop at all, Cadenya is a hosted agent loop with approvals, conversation state, resume and model swap built in, and it calls Vercel Sandbox as a tool set. A coding agent can set it up with no account, in a temporary workspace you claim later. What it offers is at the end.
Every price and limit below comes from each vendor’s own pages, checked on November 11, 2026.
Vercel Workflow vs DBOS vs Temporal vs Inngest for AI agents
Coding agents that pick Vercel Workflow usually weigh it against DBOS Transact, Temporal and Inngest. Here’s how they line up, with Cadenya last:
| Vercel Workflow | DBOS Transact | Temporal | Inngest | Cadenya | |
|---|---|---|---|---|---|
| The agent loop | Your code, in a "use workflow" function, often with WorkflowAgent | Your code, as a DBOS workflow in your app | Your code, as a workflow on your workers | Your code, in Inngest functions with steps | Written and hosted by Cadenya. Your app starts an objective over HTTP or the SDKs |
| Human in the loop | defineHook or createHook, resumed with resumeHook(), or a createWebhook() URL | DBOS.recv() waits for a DBOS.send() message | Signals and Updates | step.waitForEvent() | Mark a tool as needing approval. The loop emits toolApprovalRequested and your backend approves or denies |
| Run it yourself | Postgres World: PostgreSQL plus graphile-worker, started with your server | A library plus Postgres, with “no separate orchestration server” | Frontend, History, Matching and Worker services, plus persistence and visibility stores | One binary, SQLite and in-memory Redis by default, Postgres and Redis optional | Nothing to run. The loop runs on Cadenya |
| Managed price | Hobby includes 50,000 events a month. Pro pays $0.02 per 1,000 events, plus data, functions and queues | DBOS Pro: $99 a month with 1 million checkpoints | Temporal Cloud: Pay-As-You-Go with no base fee, from $50 per million actions | Free with 50,000 executions, Pro from $99 a month | Free for 1,000 loops a month, $49 a month for 50,000, then $0.0015 a loop |
| License | Apache-2.0 | MIT | MIT | SSPL 1.0, with an Apache 2.0 future license | SDKs for TypeScript, Python, Ruby and Go under Apache-2.0 |
DBOS and the Postgres World both keep state in Postgres, but DBOS is a library inside your process, and the Postgres World is one backend behind the same code you’d deploy to Vercel. For head-to-heads, see DBOS vs Temporal and Inngest vs Temporal.
Vercel Workflow DevKit: “use workflow”, durable steps and GA status
Vercel Workflows is generally available. Vercel announced it on April 16, 2026, and Vercel Sandbox went GA on January 30, 2026. The open-source Workflow SDK is the workflow package on npm, at 5.5.0 under Apache-2.0, and its docs live at workflow-sdk.dev (useworkflow.dev redirects there).
A workflow has two kinds of function:
- Workflow functions (
"use workflow") orchestrate. The docs say this type of function “does not have access to the Node.js runtime, and usablenpmpackages are limited,” because it has to suspend and replay. - Step functions (
"use step") do the work, with the full Node.js runtime, npm packages, and “a maximum of 3 retry attempts” by default.
Keep side effects in steps. The workflow function replays from the event log, so a model call or a database write belongs in a step, where its result is saved and reused.
Runs “are pinned to the deployment that starts them,” so an in-flight run keeps its code while new runs pick up your latest deploy. And while a run waits, “no compute resources are used until the workflow resumes execution.” The SDK has guides for Next.js, Nuxt, SvelteKit, Express, Hono and more.
Vercel Workflow DurableAgent is now WorkflowAgent
DurableAgent from @workflow/ai is deprecated. The Workflow docs put it plainly: “Use AI SDK v7’s WorkflowAgent for durable agents in Workflow 5. It replaces DurableAgent.” You import it from @ai-sdk/workflow and run it inside a workflow function:
import { WorkflowAgent, type ModelCallStreamPart } from "@ai-sdk/workflow";
import { getWritable } from "workflow";
export async function agentWorkflow(prompt: string) {
"use workflow";
const agent = new WorkflowAgent({
model: "spacexai/grok-4.6",
instructions: "You are a helpful assistant.",
});
return agent.stream({
messages: [{ role: "user", content: prompt }],
writable: getWritable<ModelCallStreamPart>(),
});
}
A tool whose execute function says "use step" becomes a durable step with retries. Durable agents with the Vercel AI SDK covers resumable streams, needsApproval and the stream transport, so this page doesn’t repeat them.
Vercel Workflow hooks and sleep for human in the loop
Hooks are how a Vercel workflow waits for a person: “workflows pause until a human takes action,” and “no compute resources are consumed while waiting.” Here’s the shape:
import { createHook } from "workflow";
export async function approvalWorkflow() {
"use workflow";
using hook = createHook<{ approved: boolean; comment: string }>();
// Workflow pauses here until data is sent
const result = await hook;
}
Your API route calls resumeHook(token, payload) from workflow/api. defineHook() adds a Zod schema, and the human-in-the-loop guide uses the tool call ID as the token.
Check who is calling before you resume. “Knowing the token does not authorize the caller,” and a createWebhook() URL is a public endpoint where “the token in that URL is the only authorization performed.” The approve button, its auth check and the screen around it are yours to build.
sleep("1d") suspends a run “without consuming any resources.” It’s a timer event in the log, not a step, so it costs two events.
Vercel Workflow self-host: the Postgres World
A World “stores workflow state and delivers queue messages.” On Vercel, the Vercel World is picked for you. Outside Vercel the default is the Local World, which writes to .workflow-data/ and is meant for development. To self-host, you switch to the Postgres World, which “uses PostgreSQL for durable storage and graphile-worker for reliable job processing”:
npm install @workflow/world-postgres
export WORKFLOW_TARGET_WORLD="@workflow/world-postgres"
export WORKFLOW_POSTGRES_URL="postgres://world:world@localhost:5432/world"
npx --package=@workflow/world-postgres bootstrap
Bootstrap creates the tables and is idempotent, so it can run on every deploy. Your server then starts the World at boot (in Next.js, from instrumentation.ts), because it runs a graphile-worker background worker.
Read the warning box before production. The docs call the Postgres World “a reference implementation” that “is not optimized for scale, speed, or security, and it ships with no authentication on the workflow HTTP routes.” They recommend cloning it and adapting it, or restricting those routes yourself.
Vercel Workflow pricing per event
Vercel bills Workflows on three meters: events, data written, and data retained. Hobby includes 50,000 events and 1 GB written a month. On Pro, events cost $0.02 per 1,000, data written $0.50 per GB, and data retained $0.50 per GB-month. Functions and queue messages bill at their usual rates.
Vercel’s own counting rule: three events for the run, three for each step that succeeds the first time, at least one more per retry, two per sleep(), and at least two per hook. So picture one agent turn as a run with 6 steps (model calls and tool calls) and one approval hook. That’s 3 + 18 + 2 = 23 events. Ten thousand turns a month is 230,000 events, or $4.60 on Pro.
At that size, the event bill is a few dollars, and the functions that run each step bill on top of it.
Vercel Sandbox: Firecracker microVMs, pricing, network policy and regions
Vercel Sandbox runs “untrusted or agent-generated code in isolated Linux microVMs.” Each sandbox gets its own Firecracker microVM “with a dedicated kernel,” and the GA post adds that code inside “is blocked from accessing environment variables, database connections, and cloud resources.” It ships a JavaScript SDK (@vercel/sandbox, 3.9.1), a Python SDK and a CLI.
Network policy is set per sandbox and changes at runtime “without restarting the process.” allow-all is the default. deny-all “Denies all outbound network access, including DNS.” A user-defined policy denies by default and allows the domains you list, and its credentials brokering adds your API key to requests for an allowed domain, so “The secrets never enter the sandbox”:
import { Sandbox } from "@vercel/sandbox";
const sandbox = await Sandbox.create({ networkPolicy: "deny-all" });
await sandbox.runCommand({
cmd: "node",
args: ["-e", 'console.log("Hello from Vercel Sandbox!")'],
stdout: process.stdout,
});
await sandbox.stop();
Vercel bills Active CPU, so CPU time spent waiting on a model call doesn’t count. In iad1, Pro pays $0.128 per Active CPU hour, $0.0212 per GB-hour of provisioned memory, and $0.60 per million sandbox creations. Vercel’s “AI code validation” example (5 minutes, 2 vCPUs, 4 GB) comes to $0.0092. Vercel Sandbox runs in 19 regions, including fra1, lhr1, sin1 and hnd1, with iad1 as the default.
Inside a workflow, the Sandbox class is serializable and its create, runCommand, stop and snapshot methods “implicitly run as steps.” Outside Vercel, the SDK authenticates with an access token. Sandboxes for running agent-generated code compares sandboxes on isolation and price, and Running agent code in E2B shows the tool set pattern that works the same way with Vercel Sandbox.
What does Cadenya offer that Vercel Workflow doesn’t?
Vercel Workflow makes your agent code durable. Cadenya is the agent loop itself, hosted, with approvals, conversation state, resume and model swap built in, and it calls Vercel Sandbox as a tool set. Here’s what that covers:
- The loop, already written. Cadenya calls the model, runs the tools, streams each reply and compacts the context window at a threshold you set. On Vercel Workflow, that’s the agent, steps and hooks you write and test.
- Nothing to deploy or pin. Your app calls an API, so there’s no workflow function to replay, no deployment pinning to plan around, and no World to pick or self-host.
- Human-in-the-loop approvals with steering. Mark a tool as needing approval, and Cadenya “parks the Agent loop until it receives approval or a denial with optional steering.” The objective emits
toolApprovalRequested, and a denial carries a memo the agent reads, such as “Use the read-only lookup tool instead.” - Conversation state and resume. Every objective keeps a durable event log of its loop. Stream events over server-sent events, reconnect with the last event ID, and send a Continue message when the objective waits for the user.
- Vercel Sandbox as a tool set. Cadenya runs the agent loop and calls your sandbox as a tool set. Give the agent a Bare
run_codetool and answer each call from a worker with@vercel/sandbox, or point an HTTP tool set at an endpoint that does. Tool sets also come from an OpenAPI 3 spec or an MCP server. - Model swap with no deploy. The model is a setting on an agent variation, not code. Use your own provider keys and pay the provider directly.
- SDKs in four languages. TypeScript (
@cadenya/cadenya), Python, Ruby and Go are all at 1.9.0 under Apache-2.0, and signed webhooks follow Standard Webhooks. - A price in one line. Free for 1,000 loops a month, $49 a month for 50,000, then $0.0015 a loop. A loop is one LLM request.
Approve or deny tools walks through the approval flow, and the free plan needs no card.
No account yet? A coding agent can create a temporary workspace with cadenya workspaces create --temporary, run its first objectives there, and hand you a link to claim it, as auth.md explains. To create an account, sign up at app.cadenya.com/signup, then copy your API key from the API keys page at app.cadenya.com/account/apiKeys.
Related articles
- Durable execution for AI agents (2026)Durable execution
- Spring AI agents in production (2026)Spring AI
- Hatchet for AI agents (2026)Durable execution
Grow wherever AI goes next.
Start shipping agents that are equipped to evolve.