On this page
Choose the layer you want to own
Choose a platform by following one agent change from editing through testing, deployment and run review. Compare the application your team needs to maintain at each step.
Use the branching-flow, run-review and regression-testing guides to inspect Cumulus. Their screenshots use synthetic data; measure cost, latency and accuracy with your own workload.
Compare responsibilities
| Approach | What it gives you | What your team still owns |
|---|---|---|
| Custom agent application | Control over your code, providers, data model and user experience | Design and operate the editor or configuration format, versioning, runtime, business integrations, tests, run history and review UI you need |
| Retell | A voice-agent platform with conversation flows, tools, published versions, environment tags and simulation testing | Your business endpoints, policies, integration acceptance and the surrounding application workflow |
| LiveKit Agents + LiveKit Cloud | A realtime agent framework, browser Agent Builder, managed deployment/scaling, session observability, workflow abstractions and testing tools | Agent/application code and the business data model, integration contracts and product experience you choose to build around it |
| Cumulus Talos | Studio plus a shared Agent/Publication/Run model, cloud and phone execution, tests, run-linked events/receipts, and registry-generated REST/SDK/MCP interfaces | Business rules, approved knowledge, credential/destination configuration, carrier setup and acceptance of your real integrations |
Retell includes versioning and simulations; LiveKit Cloud includes deployment and observability. Compare how each supports your full workflow.
If you are building it yourself
Walk one change through your stack: edit behavior, validate it, create a candidate, run regression cases, deploy it and investigate a failed run. Identify where each record lives and how you link it to the executed version.
Cumulus supplies a connected path for that work:
- Authoring: a visual flow and draft edits through the same definition.
- Release identity: immutable publications and runs pinned to the version they execute.
- Testing: suite revisions, trials, assertion evidence and candidate publications.
- Execution review: transcript, event history, policy decisions, action receipts, outputs and usage linked to the run.
- Integration surface: REST, TypeScript/Python SDKs and MCP generated from the operation registry.
With Cumulus, your team configures these features in an existing application. You still write the business endpoints, define behavior and test real integrations. Build a custom application when you need control over the editor, runtime or review experience.
If you use Retell
Retell's conversation flows, versioning and tags and simulation tests cover substantial voice-agent workflows already.
If your workflow also needs cloud tasks or chat, inspect how Cumulus connects those runs to phone agents, action receipts and the REST, SDK and MCP interfaces.
One concrete difference to test is external-action behavior during simulation. Retell’s simulation guidance says unmocked custom functions may call their real endpoints, so function mocks must be configured to avoid live writes. In Cumulus suite simulations, an unmocked customer HTTP action is refused before sending and recorded as unmocked_action; see the regression workflow. This distinction does not replace testing your real integration separately.
Cumulus supports importing a Retell export. Treat import as a starting point: validate the imported definition, inspect provider/tool/telephony mappings, run your regression suite and complete a real integration test. Import support is not proof of identical behavior or a drop-in migration for every Retell feature. See agent import.
If you use LiveKit Cloud
LiveKit provides managed deployment, workflow building blocks and testing/evaluation. Its framework is a strong fit when you want to write and operate the agent application yourself. LiveKit also offers a browser Agent Builder for prototyping and deploying agents without writing code.
Evaluate Cumulus when you want business users to work in an existing Studio and your engineering team to integrate against its agent/publication/run contracts. Compare how you would expose version changes, test evidence and action receipts to an operator in each approach. Check how each approach records business outcomes and lets operators investigate them.
Cumulus uses LiveKit technology for voice execution. Compare the agent application and operator tools separately from the underlying media service.
Evaluate with your own acceptance case
Pick one workflow with an external action and a failure path. Build it, run three representative cases, inspect a failed trial, promote a tested version and trace a real run back to that publication. Check the downstream action receipt as well as the conversation.
Record the implementation you still need, whether operators can use the evidence, and the measured quality, latency and cost under your own workload. Use the result to decide whether the platform meets your requirements.
Competitor links were reviewed on 7 October 2026. Features change; verify the linked official documentation before making a purchasing or migration decision.