New plnt v0.1 — the orchestration runtime for agentic workflows
Y Open source · Apache-2.0

Ship agentic workflows to any GPU backend.

plnt is the orchestration runtime for micro-agent workflows. Pick a workflow from the registry, pick a Kubernetes GPU cluster, and plnt handles the Helm deploy, canary, and rollback — as a durable Temporal saga.

plnt run review-responder
$ plnt run review-responder@1.2.0 --backend gpu-cluster-01
04:17:33 pull workflow=review-responder@1.2.0 · source=s3://microagents/
04:17:34 resolve backend=gpu-cluster-01 · gpus=2× available · region=us-east
04:17:34 workflow temporal://plnt/orchestrate · run=r-9c4f218e0a
04:17:35 validate spec ok · tool bindings ok · gpu class satisfied · image pullable
04:17:41 helm install release=review-responder-canary · replicas=1 · traffic=5%
04:19:14 smoke 10 sample invocations · p50=1.1s · errors=0 · pass
04:19:15 promote canary → stable · replicas=1→3 · traffic=100%
04:19:15 ready https://review-responder.plnt.svc/invoke
Runtime
Temporal saga
durable · resumable
Registry
S3 / OCI
version pinned
Backend
Any K8s + GPU
kind → prod
Deploy
Helm + canary
5% → smoke → 100%
01 / 05 — THE STACK

Three repos, one story. An end-user product that needs narrow, reliable AI features. A public registry of workflow recipes anyone can pull. A runtime that turns any Kubernetes cluster into a place those recipes can run.

architecture · top-down
    ┌────────────────────────────────────┐
    │  google-business   (end-user SaaS)   │  reviews · posts · bookings
    └───────────────────┬────────────────┘
                        │  invokes
                        ▼
    ┌────────────────────────────────────┐
    │  microagents  (workflow registry)  │  pluggable recipes on S3
    │  review-responder · post-generator │
    │  booking-triage  · trend-monitor   │
    └───────────────────┬────────────────┘
                        │  pulls spec
                        ▼
    ┌────────────────────────────────────┐
    │  plnt  (orchestration runtime)     │  ← this repo
    │  WorkflowRun CRD · Temporal saga   │
    │  Helm install · canary · rollback  │
    └───────────────────┬────────────────┘
                        │  helm install
                        ▼
    ┌────────────────────────────────────┐
    │  Kubernetes GPU backends           │  kind · GKE · EKS · on-prem
    │  scheduler · nvidia.com/gpu        │
    └────────────────────────────────────┘
02 / 05 — WHAT PLNT DOES

Six primitives. Enough to turn a workflow spec into a running service on a GPU cluster, without hand-rolled scripts.

01

Workflow-first

Point plnt at a workflow spec pulled from any OCI or S3 registry. Steps, tool bindings, GPU class, replicas — all declared, not scripted.

02

Backend-agnostic

Any Kubernetes cluster with GPU nodes. Same YAML runs on kind, GKE, EKS, on-prem. Pick the backend at deploy time.

03

Canary by default

Every run is a canary first. 5% traffic → smoke test → promote. Fails fast, rolls back cleanly. Compensation via Temporal, not bash scripts.

04

Durable saga

Every deploy is a Temporal workflow. Process restarts resume from the last completed step. Nothing hides in bash state or in-memory queues.

05

Multi-GPU aware

Workflows declare GPU class + count. plnt schedules across node pools, respects nvidia.com/gpu limits, streams GPU utilization back to the caller.

06

One CRD, one contract

A single WorkflowRun CRD is the entire deploy trigger. kubectl apply. The operator does the rest.

03 / 05 — HOW

Three steps. No control plane to run. No deploy scripts to maintain. The Kubernetes and Temporal primitives do the work.

01
Pull

Fetch a workflow spec from microagents (or any S3/OCI registry). Verified by hash, pinned by version.

02
Orchestrate

plnt starts a Temporal saga: validate → helm install canary → smoke → promote or rollback. Durable, resumable, inspectable.

03
Serve

Workflow is reachable at <name>.plnt.svc/invoke. Uniform request shape across every workflow.

04 / 05 — RUN ONE

One CRD. One Helm install. One curl.

examples/review-responder.yaml
apiVersion: plnt.work/v1
kind: WorkflowRun
metadata:
  name: review-responder
spec:
  workflow:
    ref: review-responder@1.2.0
    registry: s3://microagents
  backend:
    cluster: gpu-cluster-01
    gpuClass: nvidia.com/h100
    gpuCount: 2
  replicas:
    min: 1
    max: 4
  canary:
    trafficPercent: 5
    smokeTest:
      invocations: 10
      p95BudgetMs: 2500
plnt · deploy from kubectl
# apply the WorkflowRun resource
$ kubectl apply -f review-responder.yaml
# plnt operator picks it up and starts the saga
$ plnt logs review-responder --follow
# invoke it
$ curl -s https://review-responder.plnt.svc/invoke \
    -H "content-type: application/json" \
    -d '{"review":{"rating":2,"text":"..."}}'
05 / 05 — GET STARTED

The playground runs live. The repos are open. The docs walk through a 60-second kind demo.