CumulusDocs
Open Studio
On this page

Choose an agent platform

Use a concrete workflow to compare implementation work: book an agreed service visit, handle a stale slot and offer backup dispatch if the primary cannot connect. Review both normal behavior and failed or uncertain external writes.

Compare responsibilities

Your application still decides who may access a record, whether a booking is allowed and what to do if a transaction result is unclear. Compare who authors and executes the conversation, how a version is pinned and what evidence an operator can inspect.

ApproachQuestions to resolve
Custom agent applicationWho builds and operates graph configuration, execution, testing, publication and review tools?
RetellWhich supported graph, integration, simulation and versioning behavior meets your exact acceptance case?
LiveKit Agents and managed hostingWhich parts belong in agent code, media infrastructure and your own operator application?
CumulusCan your team edit the conversation in Studio or TypeScript, connect its business tools, and explain a failed booking from the run history?

Compare each option using the same booking task and failure cases, so the result reflects the work your team will actually do.

If you are building it yourself

For backup dispatch, account for graph editing, immutable releases, unavailable destinations, transfer consent, retries and operator review. Decide which interfaces you need to maintain. Cumulus lets Studio and the SDK edit the same conversation graph. Your business endpoints remain in your backend, while Cumulus runs the published conversation and records its progress.

The SDK tutorial shows the minimum working boundaries. Do not require a separate service for every tool or graph component.

If you use Retell

Inspect the import workflow, compare provider and tool mappings, and review unsupported settings before publishing a translated graph. After import, test the conversation paths, tool results and phone settings that your workflow uses.

Use the same synthetic scenarios and explicit endpoint mocks in each environment. Check unmocked-action behavior deliberately rather than assuming a simulation cannot contact your services.

If you use LiveKit Cloud

Separate media hosting from graph authoring, business effects and review. Decide which application behavior you want to own in code and which operators need to change in an editor. Cumulus uses LiveKit for voice media and adds graph authoring, publication and run-review workflows around it.

Evaluate with your own acceptance case

Run a normal booking, an unavailable slot, a lost response after a possible reservation and a failed primary transfer. Keep the same business contracts and synthetic inputs. Inspect the publication, transcript, effects and authoritative downstream result.

Compare developer effort, operator comprehension and failure recovery. Record provider, deployment, direction, observed audio and artifact states for any real call. Use release guidance to distinguish a tested candidate from the live publication and its number assignments.

Content version f8624d70Customer SDKMarkdown source
Choose an agent platform · Cumulus developer docs