Choose a deployment profile
All profiles can use
@agentium/eval and @agentium/observability. Add @agentium/admin for hosted configuration management; CLI scaffolding is development tooling, not your deployment platform.
Rehearse your deployment profile
Every application needs observable success and failure outcomes. Choose checks that exercise your actual host and effects:
For model behavior, run the quality gate and confirm that a deliberate regression fails. Its default fixture verifies evaluation plumbing. Use representative cases with your deployed model for a separate quality assessment.
For every long-lived host, stop admission, drain active work, close owned resources, and flush telemetry. Operations explains that sequence. Successful HTTP delivery, a completed model run, a passing evaluation, and a confirmed external effect are different observations.
Make state survive the right failure
For a payment, notification, or call whose remote outcome is unknown, preserve the operation identity and reconcile before retrying. Use the recovery guide and telephony callback guide for those decisions.
Operate the service
- Quality: evaluate representative cases and scorer assertions with
@agentium/eval; use conversation tests for multi-turn behavior. - Diagnostics: correlate model, tool, run, and request events with
@agentium/observability; choose OpenTelemetry, Langfuse, or a host sink. - Limits: bound concurrency, input/output, execution, retention, and retries. Use operating limits; cost estimates are not provider invoice guarantees.
- Administration: scope the admin API to authorized operators and persist configurations deliberately. Entity CRUD changes application configuration, not your infrastructure.
- Upgrades: pin a tested package set and follow v4 migration and queue metadata migration before changing persistent deployments.