Manager tools versus handoffs: establish who owns the final reply
Compare calling a sub-agent as a tool with handing over a conversation, and define responsibility for context transfer, final replies and failure recovery.
Choose an approach
Use a specialist as a tool when one agent should retain task organization and the final response. Use handoffs when a specialist should take over the subsequent conversation. They can be combined, but users should know who is handling the task. Choose based on control and context transfer rather than ranking centralized and distributed designs.
Returning partial work differs from taking over a conversation
The OpenAI Agents SDK distinguishes agents as tools from handoffs. With tools, the manager receives partial results and continues the response; a handoff transfers subsequent handling to the target agent. Passed context depends on configuration; do not assume every role sees the full history.
A final organizer can reconcile differences when separately checking two sources. Conversation transfer better fits users who need to continue clarifying with a specialist. Role labels do not establish correct handoffs; inspect the task and tools actually received.
Check context and final replies on the same request
Use a request requiring both research and explanation. Record subtask inputs and outputs for tool mode, and retained conditions when the receiving agent continues in handoff mode. Require both final conclusions to trace to the same material so user constraints survive transfer.
Add a follow-up requiring a return to the original role and check routing. A manager must not present an empty specialist result as verified. No SDK example was executed here; this defines control-flow validation without claiming tested multi-role processing.
Separate routing changes from content errors
Tool mode requires input-output contracts; handoffs additionally need transfer conditions and continuation rules. Count calls across all participating roles, not just the final response. Shared material changes also require checking that roles use the same version.
Start with a few clear roles and replayable tasks, then adjust based on failures. Repeated back-and-forth transfers suggest overlapping duties or unclear routing, not necessarily too few roles. Preserve a simple baseline to check whether added orchestration solves the original interaction problem.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.