> ## Documentation Index
> Fetch the complete documentation index at: https://docs.karta.sh/llms.txt
> Use this file to discover all available pages before exploring further.
# Production architecture review
> A review path for platform, security, and enterprise teams evaluating how Karta runs agents in production.
Use this page as the architecture review index. It ties together the agent
runtime, account-management systems, credential model, durable kartas, spend
gates, releases, and audit surfaces that matter before a production launch.
## First-pass model
Karta separates the code that runs agents from the systems that manage identity,
money, keys, billing, webhooks, and audit.
```text theme={null}
Your app or SDK
-> Karta session API
-> isolated agent runtime for one session and karta
-> typed event stream back to the caller
Karta management systems
validates credentials, enforces budgets, stores account state, records audit,
manages releases, keys, billing, BYOK, webhooks, and team access
```
The security property to review is the boundary: a running agent handles the
current session, but it does not become the account administrator, billing
system, or key manager.
## Review checklist
| Concern | What to confirm | Read |
| ---------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- |
| Agent execution boundary | Each session runs in an isolated runtime environment, and durable workspace state belongs to the selected karta. | [How Karta works](/concepts/how-karta-works), [Multi-tenancy](/concepts/multi-tenancy) |
| Memory and privacy grain | Your product chooses the stable karta id: per user, per virtual employee, per backend task, or per fleet member. | [Kartas & memory](/concepts/instances-and-memory) |
| Prompt injection containment | Hostile instructions are assumed possible; review what the platform contains and what your agent tools still own. | [Security model](/security/security-model), [Inputs](/api-reference/inputs) |
| Browser credential safety | End-user browsers never receive a `kt_live_` API key. They use a short-lived session token or a publishable widget key fenced by allowed origins. | [Session tokens](/security/session-tokens), [Chat widget security](/security/chat-widget) |
| Server credential authority | API keys are scoped and revocable. Use the narrowest scope for CLI, CI, backend, and management integrations. | [API keys](/platform/api-keys), [API authentication](/api-reference/authentication) |
| End-user identity | Anonymous, soft, and verified identity are distinct. Verified identity is the path for durable per-user memory and sensitive host-attested actions. | [Widget identity](/sdks/widget/identity), [Authenticated agent](/sdks/widget/authenticated-agent) |
| Spend control | Requests are gated before execution when credits or caps are exhausted. Caps can be set at org, key, user, karta, seat, and embed-key grains. | [Usage & budgets](/platform/usage-and-budgets), [Errors](/api-reference/errors) |
| Model-provider keys | Managed model access and BYOK are separate choices. Provider keys are encrypted at rest and fetched only when needed for a request. | [BYOK](/platform/byok), [Trust & compliance](/security/trust) |
| Release control | Deploys create immutable releases; activation and rollback are pointer flips under the same stable agent URL. | [Deploy loop](/deploy/deploy-loop), [Releases & rollback](/deploy/releases-and-rollback) |
| Auditability | Material account changes write immutable audit events. Use webhooks for events your own systems must receive. | [Audit log](/platform/audit-log), [Webhooks](/platform/webhooks) |
| API confidence | The run surface is uniform: sessions, messages, approvals, streaming events, and adapter shapes. | [API reference](/api-reference/introduction), [Streaming events](/concepts/streaming-events) |
| Failure handling | Budget exhaustion, auth failures, rate limits, stream errors, and provider failures have documented behaviors. | [Troubleshooting](/platform/troubleshooting), [Errors](/api-reference/errors) |
| Launch readiness | Access, spend, release, webhook, status, and support controls are assigned before real users arrive. | [Production readiness](/platform/production-readiness) |
## Integration patterns
Your server holds a scoped API key, creates sessions, sends messages, and streams events to your own UI.
Your page ships a publishable embed key, fenced by allowed origins, and Karta mints browser-safe session credentials.
Your backend vouches for the end user so Karta can route durable memory to the right karta.
Keep an OpenAI or Anthropic-shaped client and point it at a Karta agent adapter.
## Questions to answer before launch
* Which stable id creates the karta for each user, virtual employee, backend
task, or fleet member?
* Which client surfaces need API keys, session tokens, publishable embed keys, or
adapter credentials?
* Which spend caps should be hard stops, and who receives threshold webhooks?
* Which people can deploy, roll back, mint keys, configure BYOK, and view audit?
* Which events should your incident, billing, analytics, or CRM systems receive?
* Which user identity level is required before the agent can read durable
per-user context or request sensitive actions?
* Which release rollback and troubleshooting steps should be in your runbook?
## Read next
The request path, isolation model, and runtime/management separation.
Encryption, authentication hardening, secrets, webhooks, audit, and operator access.
Prompt injection assumptions, containment boundaries, credentials, and residual risks.
Launch checklist for releases, access, spend controls, webhooks, audit, and support.
API keys, session tokens, route families, and auth failures.
Metering, caps, credits, threshold alerts, and 402 behavior.
CSP, origin allowlists, graceful limits, privacy, and disclosure guidance.
Operational symptoms, likely causes, and the first fix to try.