The history of workflow orchestration in enterprise software is a history of increasing complexity gracefully absorbed into lower-level infrastructure. Cron jobs gave way to task queues. Task queues gave way to DAG-based orchestrators. DAG-based orchestrators gave way to event-driven platforms. Each transition happened because the workflows that enterprises needed to run became more complex than the prior generation of tooling could express correctly.
The AI era is forcing the next transition. The problem is not just complexity — it's non-determinism. AI steps in a workflow don't behave like database queries. They fail in ways that aren't binary. They produce outputs that need evaluation before the next step runs. They sometimes require a human to review a result before the workflow continues. None of this fits cleanly into the execution model that existing orchestration tools were designed around.
What durable execution means
Durable execution is a specific architectural pattern: the workflow engine persists the state of every running workflow to storage, such that if the worker process crashes mid-execution, the workflow resumes from exactly where it left off when a new worker picks it up. This is table stakes for AI workflows. An LLM call that takes 30 seconds might fail halfway through. A multi-step AI pipeline that's been running for two minutes can't simply be restarted from the beginning.
The existing infrastructure doesn't handle this well. Lambda functions are stateless by design. Message queues don't carry execution state. Most task runners assume that a task either completes or fails cleanly, not that it's in a partially complete state that needs to be resumed. Building durable execution on top of these primitives requires a significant amount of glue code — and that glue code is where most AI workflow failures happen.
The opportunity
The companies building durable execution infrastructure for the AI era are solving a problem that almost every enterprise engineering team will encounter in the next two to three years. This is not a niche. It's a horizontal infrastructure layer that sits underneath AI-native applications in every vertical. The defensibility of these products compounds with usage: the more workflow logic a team encodes into the orchestration layer, the harder it becomes to migrate off. That's the pattern we look for.
We backed Inngest because they understood this architectural shift earlier and more precisely than anyone else we'd talked to. The product design reflects a deep understanding of where existing tools break down in AI-in-the-loop scenarios — and a clear opinion about what the right abstraction looks like. That's the kind of founder conviction that makes us write a check before the market has caught up.