AI Deck

Inngest — A development platform that turns ordinary code into fault-tolerant workflows with automatic retries and state management

A development platform from Inngest Inc. (US) for writing backend asynchronous processing as workflows that are hard to break. If you divide your code into units called steps, a failure partway through does not restart everything from the beginning — execution resumes from the failed step (durable execution). The biggest advantage is that you do not have to set up queues, workers, or tables for storing job state yourself. It is used for background jobs, scheduled runs, webhook processing, and in recent years as an execution layer for AI agents. The SDKs are published as open source under Apache 2.0 and support TypeScript/Node.js, Next.js, Python, and Go. Released in October 2021.

Key Features

  • Durable execution: When you split processing with step.run(), the result of each step is persisted. If a network error or a temporary outage of an external API knocks the run over, successful steps are skipped and the run retries automatically from the point of failure. This makes it easier to avoid re-execution accidents such as double-charging a customer
  • No queue or worker management: You do not need to run Redis or SQS, keep worker processes resident, or maintain a database table holding job state. You only write the side that sends events and the functions that run in response
  • Declarative flow control: Concurrency limits, throttling, rate limiting, per-tenant fairness (so one user cannot monopolize capacity), and singleton execution that allows only one run per key can all be written as function configuration
  • Long waits and resumption: With step.waitForEvent() you can write logic such as “wait until an approval event arrives” or “resume seven days later.” Waiting does not occupy a server, which suits workflows with a human approval step and AI agents that pause and resume
  • Observability and replay: Per-run traces, step-level inputs and outputs, and replays of failed runs are available from the dashboard. In 2026 the platform added Insights, which lets you query events and runs with SQL, and AI Overview, which shows cost and performance of AI calls
  • Features for AI agents: Alongside the AgentKit agent framework, Inngest offers Scoring for attaching quality scores to runs (usable for LLM-as-judge evaluation), Sessions for grouping related runs, and Experiments for comparing step implementations on live traffic
  • Deploy anywhere: It runs on serverless platforms (Vercel, Cloudflare Workers, and others) as well as on long-running servers. With Connect you can also have containers open persistent outbound connections

Pricing

PlanMonthly priceWhat it includes
Hobby$050,000 executions/month, 5 concurrent steps, 500,000 events/month, 24-hour trace retention, basic tracing, metrics, and alerts
ProFrom $991M executions/month (pay-as-you-go up to 20M), 100+ concurrent steps ($25 per additional 25 steps), 5M+ events/month, 7-day trace retention
EnterpriseContact salesCustom limits, 90-day trace retention, SAML, RBAC, audit trails, dedicated Slack support

The Hobby plan can be started without a credit card. Pro is designed so that executions, span data, and events are billed on a pay-as-you-go basis above the included amounts, so the actual invoice depends on how many steps your workflows use.

Pricing is as of August 2026. Check the official site for the latest information.

Pros & Cons

Pros

  • Removes the work of building and operating queues, workers, and state tables, so asynchronous processing can be written as ordinary functions
  • Automatic retries from the failed step make it natural to write code that tolerates temporary outages of external APIs
  • Concurrency, rate limiting, and tenant fairness are expressed as configuration, so load-control logic does not get scattered through the application
  • The SDKs are open source under Apache 2.0, and a local dev server lets you verify behavior without connecting to the cloud
  • Long-running workflows are possible even on serverless platforms, because waiting does not consume the function execution time limit
  • The free Hobby plan allows up to 50,000 executions per month for evaluation

⚠️ Cons

  • The first paid plan starts at $99/month, which is a large step up for small projects that outgrow the free tier
  • Pay-as-you-go billing is based on executions, so consumption grows as you split work into finer steps; estimating cost takes some experience
  • You need to understand the constraints of the programming model — dividing work into steps and keeping values passed between steps serializable
  • The execution layer is a managed service, which adds an external dependency (self-hosting is an Enterprise-level discussion)
  • Trace retention on the free plan is only 24 hours, which is short for investigating incidents after the fact

Comparison with Similar Services

ItemInngestTrigger.devTemporalAWS Step Functions
How runs startMainly event-drivenMainly direct function invocationWorkflow invocationState machine execution
Workflow definitionCode (step units)Code (task units)Code (per-language SDKs)JSON / ASL definition
LanguagesTypeScript, Python, GoMainly TypeScriptGo, Java, TypeScript, Python, and othersLanguage-agnostic (connects Lambda and others)
Self-hostingOpen source SDKs plus managed executionSelf-hosting availableSelf-hosting available (OSS)Not available (AWS managed)
Free tierYes (50,000 executions/month)YesTemporal Cloud has a free tierCovered by the AWS free tier
Best suited forAsynchronous processing in SaaS, AI agentsTypeScript-centric job executionLarge-scale, long-lived core workflowsOrchestration that stays inside AWS

Who Is It For

  • Backend developers tired of assembling their own job queues on Redis or SQS and looking after retries and dead-letter handling
  • Anyone who wants to write long-running processing on Vercel or Cloudflare Workers that does not fit within the execution time limit
  • Teams that call several external APIs in sequence (payments, email delivery, data sync) and want careful handling when a step fails partway
  • Developers running AI agents in production who need an execution layer with pausing, resumption, retries, and evaluation logs
  • Operators of multi-tenant SaaS who want to manage concurrency and rate limits declaratively so that one tenant cannot occupy all capacity

Summary

Inngest takes over the queue, worker, retry, and state-management plumbing that teams rebuild every time they write asynchronous processing. Once you learn the habit of splitting code into steps, failure-tolerant long-running workflows can be written as straightforward functions. In recent years its operational features for AI agents, such as Scoring and Sessions, have grown, giving it a stronger character as an agent execution layer. A clear first move is to take one existing background job and move it over using the free Hobby plan and the local dev server.

← Blog