Why Long-Running Agents Break
Agents that run for minutes or hours hold state, wait on external tools and compete for shared resources. Traditional request-response infrastructure was never designed for that shape of work.
Retries create duplicate actions, crashed workers lose progress, and two agents editing the same record can silently overwrite each other.
The Ephemeral State Machine Model
Each agent run is modelled as an explicit state machine with typed transitions, so every step is resumable.
State is checkpointed after every tool call, which makes retries idempotent instead of dangerous.
Short-lived leases on shared records prevent concurrent writes without global locks.
“Concurrency bugs are not solved by faster workers; they are solved by explicit state.”
Running at Scale
With ephemeral state machines, NOVA schedules long-running agents across its worker fleet while guaranteeing that each action happens exactly once.
