FinOps wedge · Money on the table · PAYG agents

Find idle spend. Ship the YAML.
Then let gated agents heal what burns.

First session: surface wasters and rightsizing actions — not another suite demo. Observe. Heal. Learn. Specialized PAYG agents (Observer → Actor → Critic → Executor). We dogfood Avirka-EN on our own stack. No suite tax.

Vanilla
Open standards · no suite tax
Specialized
Agents you own & require
Sovereign
Your VPC · your weights
PAYG
Idle → $0 · pay as you go

Built for modern platform & SRE stacks

KubernetesGKEVertex AIvLLMQdrantPostgresOpenTelemetryPrometheusKnativeHugging FaceArize PhoenixGrafanaKubernetesGKEVertex AIvLLMQdrantPostgresOpenTelemetryPrometheusKnativeHugging FaceArize PhoenixGrafana
Conversion wedge · First session ROI

Money on the table — before the agent theater.

Land with FinOps wasters and a concrete remediation path. Expand into shadow Reflexion, then policy-gated healing when trust is earned. Scenario framing only — we don't invent customer dollar claims.

Idle node pools

rightsizing YAML before the next renewal

GPU headroom

condition on p99, not mean util

Suite tax

hosts × seats × SKUs vs decision volume

First-session path

  1. Connect billing + cluster signals (BYO when ready).
  2. See wasters list — idle pools, GPU headroom, cost spikes after deploys.
  3. Copy the rightsizing / remediation YAML; Actor proposes, Critic gates.
  4. Raise autonomy only after measured false-approve rates — shadow default.

Suite AIOps vs Warble shape

Category comparison — not a vendor scoreboard. Patterns first; product second.

DimensionWarbleTypical suite AIOps
What you buySpecialized agent SKUs · pick & chooseBundle / seat tax you rarely use
Bill shapePAYG meters (incident · proposal · eval · execute)Hosts × seats × feature SKUs
Safety modelIndependent Critic + OPA · shadow defaultChat widget bolted on the APM
Where data livesYour VPC · sovereign path availableVendor cloud traces by default
Time to valueFinOps wasters list → YAML in first sessionDemo polish → renewal step-function

Named competitors (Datadog Bits AI, incident.io AI, PagerDuty AIOps) sell depth of integration; we sell bill shape, sovereignty, and gated Actor/Critic — pick what fits your renewal math.

Design partners

We don't invent customer logos. First three design partners get this slot — until then, dogfood is our proof: we run Avirka-EN on our own platform.

Ask about design partner seats

Pricing shape

Workspace is pay as you go per agent meter. Self-serve Stripe checkout tracks in #143.

Observe
Free tier meters

Eyes on signals. Demo-cast free units / month. No execute.

Team / PAYG
Meter what runs

obs · actor · critic · executor list rates. Shadow execute = $0.

Enterprise
Custom MSA

Sovereign VPC, committed use, SLA, private fine-tunes.

Continual learning loop

Trace incidents. Gate fixes. Improve the model.

Same vision as modern agent platforms — but purpose-built for Kubernetes SRE, dual-brain safety, and a sovereign dogfood engine you can host yourself.

What is my cluster doing right now?

Prometheus, pod events, and boutique signals feed the Reflexion engine. Accipiter watches Pending workloads. Observe is in-cluster — no black-box SaaS hop required.

avirka · observe · live path
sourceGMP / k8s events
detectormetrics-detector
scaleraccipiter-striatus
modeshadow-safe
Flock army · Autonomy · Pick & choose · PAYG

Your agents don't chat. They form up.

Drop a platform issue — the demo cast forms an army under your autonomy rules. Then procure only who ran. Pure vanilla SKUs. No suite tax. Pay as you go until memberships open.

Agent army · live formation

Drop a platform issue. Watch the demo cast form up under your autonomy rules.

Restart storm → Observer opens incident · Actor proposes limit raise · Critic gates · Executor applies if allowed.

Full army engaged · PAYG only meters what actually runs

Ruffle

Ruffle

Watches first. Warble Army forms up under your autonomy.

Core cast

Starling — Observer

Observer

Starling

Signals → incidents.

Hummingbird — Actor

Actor

Hummingbird

Draft remediations.

Owl — Critic

Critic

Owl

SLO & blast-radius gate.

Harrier Hawk — Executor

Executor

Harrier Hawk

Apply after approve.

Wings

Raven — Knowledge

Knowledge

Raven

RAG, runbooks, memory.

Sparrow — FinOps

FinOps

Sparrow

Spend, rightsizing, cost governance.

Accipiter Hawk — Accipiter

Accipiter

Accipiter Hawk

Predictive scaling.

Autonomy mode — how far agents may go

Demo cast (SKU 1–4) — pay as you go

Observer → Actor → Critic → Executor. No suite tax. Meter only what runs.

Optional add-ons (also pay as you go)

Your procure pack

Autonomy: Supervised· pay as you go

Critic + human approval required before every execute.

  • 1. Observer(obs)
  • 2. Actor(actor)
  • 3. Critic(critic)
  • 4. Executor(executor)

Illustrative mid-month list (demo volumes): $12.60 before free tier.

workspace_settings.yaml
autonomysupervised
pricingpay_as_you_go
agents:
- obs # 1/4on
- actor # 2/4on
- critic # 3/4on
- executor # 4/4on
meters:
per_incident$0.01
per_proposal$0.05
per_eval$0.01
per_execute$0.10

Default ship: supervised + demo cast obs · actor · critic · executor (SKU 1–4), all pay as you go. Knowledge, FinOps & Accipiter are optional add-ons. Shadow/dry-run never bills executor.

Closed beta · Memberships opening

Join the PAYG waitlist

Tell us which flock you want and how far they may go. No suite tax — we only open seats for packs you actually intend to run. Shadow default until you promote.

Procure pack
Preferred autonomy

PAYG · no suite tax · we never invent customer logos. Enterprise: [email protected]

Reflexion engine in action

Critic evaluates · Hypothesis proposes · Actor executes

Closed beta — sign in to try the live demo

The interactive Reflexion terminal is available to beta operators. Sign in with your beta invite, or contact us for enterprise access.

Avirka-EN · dogfood engine

We run the SRE LLM on our own platform first.

Avirka-EN is Warble's self-hosted remediation brain: Actor proposes, Critic gates, outcomes retrain. Goldens, eval-ab, train gates (≥0.70), and HF→vLLM serving close a loop most "agent demos" never ship. That's continuous learning for ops — not a chat widget bolted onto Grafana.

observesignals in
healactor ↔ critic
learneval + fine-tune
servevLLM / HF
scaleaccipiter hawk
modeshadow default
How it works

One loop runs your cluster. Another keeps training it.

Reflexion subagents — Observer, Actor, Critic, FinOps — act on live signals through Avirka-EN. Every outcome feeds continuous improvement.

Runtime · Reflexion loop
Signals
Prometheus · Loki · Alertmanager
Observer
k8s state · metrics · logs
Actor
proposes remediation
Critic
SLO + policy gate
Executor
GitOps PR · kubectl
Avirka-EN inference (Ollama · vLLM · Vertex fallback) Qdrant memory · Postgres audit FinOps agent
Training · continuous fine-tune
Incident traces
outcomes · runbooks
Airflow ingest
ShareGPT corpus
KubeRay QLoRA
fine-tune on Gemma
MLflow registry
eval-gated promote
Avirka-EN
serves the loop ↑

Every remediation outcome becomes a training signal — the model that runs your cluster is fine-tuned on your cluster.

Runs on the FRAKMA stack

Avirka-EN · Ollama
Inference
Vertex AI · Gemini
Inference
KServe · vLLM
Inference
Qdrant
Memory
PostgreSQL
Memory
Redis
Memory
Apache Airflow
Training
MLflow
Training
Argo CD
Platform
Kubernetes
Platform
Prometheus
Observability
Grafana
Observability

Why teams choose Warble

Don't ship vibes. Ship observe → heal → learn with gates you can audit.

Action Brain

Multi-step Kubernetes remediations via Critic → Hypothesis → Actor. Safe by default — policy-gated before anything reaches your cluster.

Knowledge Brain

Semantic retrieval over runbooks, incident history, and live metrics. Sub-10ms vector hits via Qdrant — no token bloat under load.

Policy-gated safety

OPA gates, pre-computed rollbacks, Workload Identity, air-gapped perimeters. Sovereign AIOps — your data stays in your VPC.

Intelligent FinOps

Correlate cost spikes with deploys, right-size pools, kill idle waste — governed by spend policy, not vibes.

Avirka dogfood engine

Avirka-EN is our SRE LLM we run on ourselves: remediates, learns from outcomes, and continuously improves via eval gates and fine-tunes.

Reflexion loops

Up to 8 self-correction cycles. The Critic stops confident-but-wrong actions before they ship — the category that kills agent trust.

Air-gapped perimeters · Workload Identity · Zero standing privileges · OPA policy gates
Security architecture

Don't ship vibes. Don't burn bills on suites.

Pure vanilla specialized agents. Avirka-EN dogfood learning. Pay as you go on your cluster.