AI customer agents, shipped like software

Agents that act.
Safely.

Kaarna is the developer-first platform for AI customer agents that don’t just answer — they process the exchange, update the subscription, issue the refund. Every message and every action is checked by independent supervisors before it reaches your customers or your systems.

The platform

What the platform does

⌨️

Journeys as code

Define goals, guardrails, and composable skills in a declarative spec that lives in git — versioned, diffed, reviewed like everything else you ship.

🧭

Any model, your choice

A model gateway routes each task to the right model — yours to configure, per task, per vendor. Bring your own keys. Automatic failover, transparent cost per conversation.

🛡️

Supervised actions

Independent supervisors check every message for policy and grounding before it sends. Irreversible actions are gated mechanically — in code, never by prompt.

🔁

Simulation & CI

Replay real conversations against every change. Persona-based stress tests. Your agent can’t regress without your build going red.

The spec

An agent you can read in one screen

Declare what the agent must achieve and what it must never do. The runtime works out the path, so you never maintain a dialogue tree.

# agent.yaml
agent: acme-support
goals:
  - Resolve order issues without handoff when policy allows
guardrails:
  - id: refund-cap
    rule: Refunds over $200 require human approval
    severity: escalate
skills:
  - use: triage
  - use: answer        # every claim cited, supervisor-verified
  - ref: ./skills/order-exchange.ts
tools:
  - shopify.lookup_order
  - shopify.create_exchange   # irreversible → confirmed & gated
tuning:
  order_exchange: { determinism: high }
Why Kaarna

Built for the teams the enterprise vendors won’t serve

Self-serve, not sales-led

Deploy a production agent on your own terms. No discovery calls, no mandatory services engagement, no opaque contract.

  • TypeScript SDK, local dev server, one-command deploy
  • Works in your existing CI and dev workflow
  • 40-line spec to first conversation

Trust you can audit

Every decision your agent makes — every plan, model call, tool execution, and supervisor verdict — is recorded in a replayable trace your compliance team can actually read.

  • Per-decision audit trail, exportable
  • Grounded answers: every factual claim cites its source
  • Fail-closed supervision on irreversible actions

Become a design partner

We’re onboarding a small group of design partners in e-commerce and SaaS. You get hands-on setup help and early pricing; we get your feedback.

Become a design partner