Agent Networks with HITL & sub-agent tool results
I've had two main issues when working with agent networks:
(1) Tool results:
When I do an agent network, the main agent does not see the sub-agent tool results. Is this intended? If so, how is the orchestrator agent supposed to know if the sub-agent executed the tool successfully or not?
I am finding it hard for the main agents to know if the sub-agent tools executed successfully or not.
For example: I tell my main agent: create a ticket... The main agent orchestrates it to the sub-agent, and even though the sub-agent creates the ticket by calling the tool, the main agent says it wasn't able to do so. (I assume because the sub-agent is not passing the tool result through)...
(2) Tool HITL:
Some of my tools have HITL enabled (requireApproval: true)... However, if I have an orchestrator agent with no tools and a sub-agent with a tool that requires approval, then I see a lot of weirdness.
- Not only is the approval request not even made... But also the orchestrator agent thinks that the tool was executed successfully when it wasn't.
Why is this? Am I doing something wrong?
(1) Tool results:
When I do an agent network, the main agent does not see the sub-agent tool results. Is this intended? If so, how is the orchestrator agent supposed to know if the sub-agent executed the tool successfully or not?
I am finding it hard for the main agents to know if the sub-agent tools executed successfully or not.
For example: I tell my main agent: create a ticket... The main agent orchestrates it to the sub-agent, and even though the sub-agent creates the ticket by calling the tool, the main agent says it wasn't able to do so. (I assume because the sub-agent is not passing the tool result through)...
(2) Tool HITL:
Some of my tools have HITL enabled (requireApproval: true)... However, if I have an orchestrator agent with no tools and a sub-agent with a tool that requires approval, then I see a lot of weirdness.
- Not only is the approval request not even made... But also the orchestrator agent thinks that the tool was executed successfully when it wasn't.
Why is this? Am I doing something wrong?
