Skip to content
RUNTIMEPersistent Agentic Runtime

The Runtime WhereAgents Actually Live

PAR separates stateless compute from durable state so your agents can pause for days, branch into hundreds of tasks, and resume instantly—without you owning any infrastructure.

Typical Latency
200–400ms
Vector retrieval + model generation
Throughput
5k reads / 1k writes/sec
Per project in production workloads
Availability
99.9% SLA
Automatic failover built in

Persistent Agentic RuntimeLive

State Layer
PAR Durable Store
MemoryCheckpointsVariables
User MessageIngest
Stateless WorkerElastic
PAR State LayerDurable

Separate Compute from State, Unlock New Agent Patterns

PAR gives every agent a durable state layer and a pool of elastic stateless workers. That simple separation is what makes long-lived, massively parallel, and human-in-the-loop flows possible.

Request Lifecycle in PAR

  1. 1
    User message or event arrives.

    A chat message, webhook, or scheduled event hits the PAR ingress with an agent and flow reference.

  2. 2
    Stateless worker is assigned.

    PAR routes the request to an available worker that spins up in milliseconds—no warm-up or manual scaling required.

  3. 3
    State is loaded from the PAR layer.

    The worker pulls variables, and checkpoints from the durable state layer.

  4. 4
    Flow executes, tools and models are called.

    Vector Flow’s execution engine runs the steps: routing, retrieval, model calls, tool invocations, and side effects.

  5. 5
    State is checkpointed and worker is released.

    Updated state is written back to PAR, a response is streamed to the caller, and the worker returns to the pool.

Built for production from day one

Automatic failover. Requests are routed across availability zones with retries and backpressure built in.

Transparent scaling. Handle spikes from 10 to 10,000 concurrent conversations without touching your architecture.

Predictable costs. Pay per operation and state, not for idle pods, over-provisioned clusters, or homegrown schedulers.

When PAR changes what's possible

Use PAR when you need agents that:

  • Pause for hours or days waiting on human approvals or external events.
  • Fan out into hundreds of parallel tasks that share and update a common state safely.
  • Maintain conversational and operational context across long-lived customer or case lifecycles.

Agent Patterns That PAR Makes Easy

Instead of fighting your infrastructure to support advanced behavior, PAR gives you these patterns out of the box.

Long-Lived Conversations

Agents that remember a customer's history for months and can re-open threads without warm-up hacks.

  • • No custom session stores required
  • • Checkpoints for critical decision moments
  • • Smart history search and retrieval built in

Massive Parallelization

Research, QA, or operations agents that fork into hundreds of concurrent tasks and merge results safely.

  • • Shared parent state with child flow inheritance
  • • Automatic fan-out and aggregation
  • • No manual worker coordination

Human-in-the-Loop Systems

Agents that propose actions, wait on approvals, and then complete the workflow once a human signs off.

  • • Durable tasks that survive restarts and time
  • • Clear audit trail via Vector Flow runs
  • • Safe execution configured per flow

Design in Vector Flow. Run on PAR. Ship real agents.

Use Vector Flow to visually design agent behavior, then rely on PAR to keep it fast, durable, and observable in production.