Multi-agent voice workflows: when to split responsibilities

Splitting a voice experience into specialists can make responsibilities clearer, but it also creates state and handoff work. Start with the smallest structure that supports the task. Several model names or nodes do not automatically produce a better caller experience.

Distinguish the boundaries

A workflow stage changes what the same assistant should do next. Backend delegation asks another component to perform reasoning or an action while the caller-facing session continues. An assistant handoff changes the active specialist. A telephone transfer moves the caller to another endpoint. These are different operations with different acceptance requirements.

Burki supports an explicit subset of workflow behavior for each runtime. GPT Live client/server delegation modes should not be treated as a promise that every historical assistant-orchestration node or provider combination runs unchanged.

Split only for a clear responsibility

A useful split might isolate collecting request details from an authorized backend lookup. Document what information crosses the boundary, who owns the result and which component may speak the final answer. Keep credentials and permissions in the appropriate service rather than embedding them in prompts.

When a caller corrects a detail while backend work is running, the old response may no longer be appropriate. Discarding a stale spoken response does not undo an external action that already completed. The application still needs to reconcile the action result.

Define state and failure handling

For each transition, specify the exact input, returned result, timeout behavior and permitted next step. Do not assume unlimited conversation context or that every specialist receives every previous message. Test missing information and contradictory state explicitly.

Repeated transitions can produce loops, duplicate questions or duplicate actions. Use a fixed scenario that crosses the boundary, returns to an earlier topic and corrects a detail mid-operation.

Verify the selected revision

Practice the exact unpublished workflow version before publishing. Unsupported nodes must be corrected; a visual graph is not proof of runtime compatibility. Inspect real transitions and tool results rather than relying on an animated diagram alone.

For human transfer, add controlled telephone acceptance separately. See the routing guide. Add more specialists only when a tested boundary improves the task enough to justify the extra operational work.