Inngest

Inngest shines when your workflows involve LLM calls, multi-step retries, and per-tenant concurrency. The durable step model is straightforward: you wrap a function call with step.run(), and Inngest automatically persists state, retries on failure, and provides tracing. The step.waitForEvent() primitive is a game-changer for AI agents that need human-in-the-loop pauses — you can wait for a user confirmation or an external event without managing webhooks or state yourself. Scoring (inngest.score()) is a differentiator: it lets you attach quality signals to runs, steps, or experiment variants, enabling LLM-as-a-judge evals and product outcome tracking at the execution layer. Combined with Experiments (group.experiment()), you can A/B test step variants against live traffic, judged on timing, retries, failures, and cost — no extra pipeline needed. The AI Overview dashboard, added in August 2026, turns the OpenTelemetry-derived gen_ai.* metadata into a live view of usage, cost, and performance across OpenAI, Anthropic, Google, and Vercel AI SDKs. This is a real productivity booster for teams that want to track model spend and latency without building custom dashboards. Where Inngest falls short: it's code-first, so if you prefer visual workflow builders, you'll be disappointed. There's no built-in visual editor for defining workflows — you write code and use the UI for monitoring. Also, while it's open-source and self-hostable, the managed cloud offers the full observability and turnkey features; self-hosting loses some of that. For strict message ordering and exactly-once guarantees, Inngest is not a fit — it's at-least-once. For simple cron jobs, Inngest is overkill — Vercel Cron or similar is cheaper and simpler. But for production-grade durable workflows with AI agents, it's a strong, cost-effective alternative to Temporal (which requires a separate workflow engine) or Celery (which requires a broker like Redis).