How to make subagents more assertive?
How can I make subagents more assertive about reporting back to the parent? I don’t fully understand the correct way for a subagent to re-enter the parent’s message queue once its job is done. Right now, they seem to lag significantly.
Here’s the test scenario I’m using, it’s just a simple models health check :
I tell the main agent:
In practice, this is what happens:
- The main agent correctly spawns the subagents
- The subagents execute properly on-time (I can see successful responses in LLM proxy)
- However, the main agent reports:
In the gateway WebUI, I can see that subagent messages only occasionally make it back into the main agent’s queue, with a delay of 1-2 minutes. By the time they arrive, the main agent has already reported that the model failed.
Then, after the main agent finishes reporting that all models were silent, a large batch of subagent responses appears in the session, arriving one after another with 1-2 second gaps. It looks like they were queued somewhere and then flushed in bulk.
I’ve tried tuning the queue behavior using
Also, most subagents in this type of task don’t return anything at all.
However, if I ask them to write their results to a file instead, they work perfectly and much faster than 13s.
At one point I even considered implementing a custom result-announcement mechanism for subagents, just to get more consistent and reliable behavior
Here’s the test scenario I’m using, it’s just a simple models health check :
I tell the main agent:
In practice, this is what happens:
- The main agent correctly spawns the subagents
- The subagents execute properly on-time (I can see successful responses in LLM proxy)
- However, the main agent reports:
"model_name" did not return a resultIn the gateway WebUI, I can see that subagent messages only occasionally make it back into the main agent’s queue, with a delay of 1-2 minutes. By the time they arrive, the main agent has already reported that the model failed.
Then, after the main agent finishes reporting that all models were silent, a large batch of subagent responses appears in the session, arriving one after another with 1-2 second gaps. It looks like they were queued somewhere and then flushed in bulk.
I’ve tried tuning the queue behavior using
/queue steer/queue interruptAlso, most subagents in this type of task don’t return anything at all.
However, if I ask them to write their results to a file instead, they work perfectly and much faster than 13s.
At one point I even considered implementing a custom result-announcement mechanism for subagents, just to get more consistent and reliable behavior
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
134,827Members
Resources
Similar Threads
Was this page helpful?
