Since the “upgrade” to multi_agent_v2, the user no longer has an ability to talk to the subagents directly.
This massively breaks the flow of a single orchestrator thread managing the feature sub-workers, starting the workers and then integrating their work, while the user refines and corrects the sub-workers understanding/implementation direction. The orchestrator thread meanwhile remains on high-level vantage point, keeping itself available to the user, and its context clean.
The flow without ability to write to subagents necessitates forwarding messages to each separate sub-worker through the parent thread, forcing extra hops, waste of the parents context on unnecessary message routing, waste of time, problems of communicating through the third-party that doesn’t have the full context of the task in question.
Such as they can’t even summarize what I precisely want to the sub-worker and return their answer to me as-is. “Verbatim” is now my most commonly used word.
I don’t understand what improvement the “v2” subagents achieve through this deprecation, compared to v1. As in, I don’t understand the justification to *remove* such a basic feature and force the user to jump through hoops.
It’s not even possible to return to v1 subagents cleanly: hacks like setting multi_agent_v2 = false or editing the model catalog work about 50% of the time and break randomly (randomly switching my already spawned subagents from normal interactive ones to “v2”). Per Codex’s analysis, part of this behavior seems to be due to (IMO inconsiderate) product decisions, and part due to just buggy implementation.
Is it just an oversight? Or is there some bigger picture, strategic direction where I soon won’t be allowed to post to the top-level thread as well?