A practical acceptance checklist for an agent pilot
Choose one job and an owner
Start with encoder guidance, playback triage or issue intake. Name the source owner and the operator responsible for unresolved requests. Select deterministic flow, evidence excerpts or model-backed conversation for that task. Write down exclusions so users know when to ask for help. The pilot should answer a specific operational question rather than demonstrate every screen.
Build representative cases
Use ordinary wording, supported questions, ambiguity and unsupported topics. Evidence cases need expected and forbidden source IDs plus abstention checks. Flow scenarios should cover branches and fallback. Keep a few questions outside the saved suite for manual review. Automated gates expose regressions but cannot establish that advice is appropriate in every customer environment.
Rehearse failure and recovery
Stop the configured model, interrupt a reply, deny microphone access and request an operator. Verify visible errors and preserved state. Restart and inspect tickets, profiles and source snapshots. Restore a backup in an isolated directory. Record the environment actually tested: a local exercise does not demonstrate production-scale load, physical-device quality or recovery on another host.
Account for costs and decide explicitly
Revenue previews quote possible plans and count completed operations; they do not charge cards or establish authoritative voice minutes. Budget inference, storage, hardware and staffing. The Media Server subscription does not set Agents pricing. Review failed cases, operator workload, audible latency and rollback evidence before expanding a pilot, and agree the commercial scope with creekX.
Product and guides
Explore creekX Agents · Quick start · Documentation. Written for v0.10 early access, 23 October 2026.
Frequently asked questions
Where should I start?
Read the separate Agents quick start and evaluate v0.10 in a private local workspace before a customer pilot.