One run is an actor. The whole fleet is a cluster.
Deterministic execution on a clustered actor runtime: every run resumes where it stopped, and resumes identically.
The operation is the load test.
A large contact center is tens of thousands of simultaneous conversations, each with tools, rules, a cost budget and a human who can take over mid-sentence. It is a distributed systems problem before it is an AI problem.
Properties of the runtime, not code per journey.
A conversation is an actor
Not a process, not a queue entry: every run is an isolated, supervised, event-sourced entity, the model that runs telephone exchanges and trading desks.
A fleet is a cluster
Entities spread across nodes on their own. Supervision, backpressure, passivation and rebalancing are properties of the runtime, not code written per journey.
A thousand children, one clock
With fan-out delegation, the parent waits for the slowest child, not the sum of all of them. Concurrency without a line of concurrency code.
Pause without losing the thread
A conversation can wait days for a document, a callback or an approval, survive a node going down and resume exactly where it stopped.
The execution model.
01
Supersteps: every node runs, a barrier applies the state, and the run can be replayed exactly.
02
Every run is a durable entity, distributed across the cluster.
03
Two layers of durability on the turn path: the conversation journal and a checkpoint per step.
04
Waits declared on external events, delivered exactly once and resilient to failover. Durable timers, bounded loops and scale-to-zero.
05
Every run resumes where it stopped. Resuming identically is what lets a decision be reproduced months later.