Description
Feature Request: “Mixture of Tool-Calling and Handoff”
Is your feature request related to a problem? Please describe.
I have a setup with 1 main agent and multiple tools (which can be other agents, or deterministic agent workflows)
Note: The teacher agent situation below is just an example. I think there are many use cases, where there should be one main agent, aggregating information from multiple tools (some of them which could be other specialized agents and others search-tools etc) where these tools have access to the full previous history.
-
Main agent – runs on a light-weight/cheap model, drives the dialogue.
-
Teacher agent – runs on
o3, aggregates external knowledge, does heavy reasoning.
I’d like to expose the entire conversation context (all user/assistant messages and prior tool results generated in the current run) to the Teacher while still treating it as a function-tool call so that control returns to the main agent afterwards.
Today I have two partial options:
| Option | Drawback |
|---|---|
| Handoff | Teacher gets full history ✔️ …but it takes over the thread, so the main agent never speaks again. |
| Function tool | Main agent continues ✔️ …yet the Teacher only sees the synthetic input the first agent passes, forcing my small model to re-summarise / concatenate everything (token heavy, lossy). |
I think it would be great to have something in between where tools (which can be agents, something different or deterministic workflows of agents) have access to the current history of the "main" agent.
Thanks for considering!