Korva is a scroll of the internet’s best ideas.
Months to build, a day with steps
The patient flows
A patient fills out a treatment form. A doctor approves it. A patient asks for a prescription. Each of those events kicked off a series of steps, and the system also had to check whether specific things had happened along the way, such as whether a doctor approved a prescription within 24 hours. If not, someone had to follow up on that request.
Months of work
Building those flows was among the worst projects the engineering team there faced. It took months for what were really basic flows, and an event-driven record suited the work because it leaves a good audit trail.
Where Inngest came from
Out of that experience came the core ideas behind Inngest: an event comes in and runs a step function, the function pauses until something happens in the product, and then it resumes automatically or times out so that it is clear the thing never happened. Those same step loops later turned out to fit AI, which calls an LLM, acts on the response, and repeats until the goal is met.
Why do timeouts matter in waiting workflows?
A workflow that waits for an event has to tell "not yet" apart from "never." A timeout turns silence into an explicit outcome that code can act on, such as sending a reminder or reassigning the task. Without one, a stalled request sits forever with no owner. Support ticketing systems use the same idea with service level agreement clocks, which commonly set a response target measured in hours.
Why does an event record give a good audit trail?
An event log is append-only: each entry says what happened and when, and old entries are never rewritten. Storing a system's history as a sequence of events instead of only its current state is called event sourcing, and it lets anyone reconstruct how a record reached its present form. Accounting ledgers work the same way, where corrections are added as new entries rather than erasures.
How do AI agent loops resemble step functions?
Each model call and each tool call can be recorded as a step, and the loop repeats until a goal condition is met. If one step fails, only that step is retried, not the whole run. The ReAct paper, published by Yao and colleagues in 2022, described this pattern of alternating reasoning and actions for language models.
More from The Peel with Turner Novak
