Question 1

Why a whole language instead of an agent framework?

Accepted Answer

The claim is that stateful loops, belief updates, DAG scheduling, and tiered memory need runtime support: actors, mailboxes, checkpoints, and schema-constrained inference are VM concepts here, not library conventions.

Question 2

How do generate and judge steps stay separate?

Accepted Answer

infer synthesizes typed values while decide evaluates explicit questions against state as a distinct durable effect, so each can use a different model class without changing control flow.

Question 3

What happens when a tool call crashes midway?

Accepted Answer

The VM checkpoints before execution and records outcomes atomically, resuming from the journal; after commit it will not re-invoke, but a service that accepted work before the crash may still see a repeat.

Question 4

Can I try decisions without paying for calls?

Accepted Answer

Yes: the triage example runs under a mock provider flag, and removing that override with a TypeSafe key set routes the same program to hosted Jev.

Question 5

How do actors communicate, and what does durability actually guarantee?

Accepted Answer

Actors are started with spawn_link and talk through mailboxes, with exit signals between them and immutable values across them, and every effect is written to a durable journal. Delivery is at-least-once, not exactly-once: an effect that crashed before it committed can be replayed, which is why any non-idempotent tool call has to carry an idempotency key you supply.

Question 6

How does measured confidence become a control-flow decision?

Accepted Answer

infer synthesizes typed values and the confidence attached to the inference becomes part of them, so a low-confidence outcome arrives as an Uncertain value that your code can branch on directly rather than a number you threshold after the fact. decide is the separate path that evaluates explicit Choice, Score and Noul questions through a provider, with a TypeSafe Jev driver bundled alongside the others.

Question 7

What should I check before pointing this at real data?

Accepted Answer

Two things the catalog record flags. Effect payloads may carry prompts or customer data, so anything you pass through infer or decide travels with the journal and needs handling accordingly. And because crashed effects can replay, non-idempotent tool calls need idempotency keys. The repository readme could not be fetched for this entry, so verify both against the source before relying on them.

Turn

Highlights

Questions about this entry

More Jev resources like this

From the Jev community