Questions about Multi-agent collab: sub-agent interruptions + orchestrator prompting + evals

[nikhil-pandey]

Problem

In Codex multi-agent/collab, the orchestrator frequently interrupts sub-agents and the sub-agents end up getting little done. Net effect: added coordination overhead, worse wall-clock completion.

Questions

1. Are there plans to improve the orchestrator prompts/default behavior so sub-agents aren’t repeatedly derailed mid-task?
2. What are the intended semantics for the messages—are sub-agents being hard-stopped vs steered?
3. Any current or planned evals comparing multi-agent vs single-agent on success rate, quality, wall-clock time, and overhead?

[JonathanFly]

Yeah Codex seems confused how it's supposed to use and manage the sub agents and their capabilities. For example, I noticed it wasn't sure how multi agent sandboxes work in the thinking traces. Maybe it's just being extra thorough by running tests, but shouldn't it know this out the box?

I need to check if the child agents inside the collaboration environment have their own sandbox as well. It seems likely that they do, but I want to confirm this through testing. I’ll prepare a prompt file to help with the test.

I also noticed Codex isn't well adapted to the asynchronous workflow. If the sub agents tasks are even moderately related or interconnected, it ends up doing things in strange orders based on what and when the sub agents happen to finish things. I think for now it'd be better to set a hard rule for the primary agent for now - just WAIT until all the sub agents are finished before moving on.

[MannyKayy]

it would also be good to get progress updates from sub agents occasionally, and maybe have a feature to be able to bring up a sub agent and not only view it, but interact with it directly.. in order to correctly guide it or steer it on intricate tasks

[ThinkOffApp]

The interruption problem described here goes away if you switch from synchronous to asynchronous coordination.

We coordinate 9 agents through shared chat rooms (IDE Agent Kit + Ant Farm). Agents post updates to rooms, a bridge relays webhooks, and the orchestrator polls for new messages with an interval from a few seconds to a couple of minutes. Nobody interrupts anybody. Agents work independently and report results when done.

The tradeoff is latency (polling adds seconds or 2 minutes depending on interval). But for most multi-agent workflows, that latency is fine because the agents are doing real work that takes minutes anyway. The reliability gain from not having synchronous signal dependencies is worth it.

[aleksandarperak33-hub]

Typed agent-to-agent messaging is the exact gap we've been working on. In TokenRip (tokenrip.com) we ended up with 7 verbs: propose, accept, reject, counter, inform, request, confirm. Each has a typed payload, signed envelope, thread-scoped delivery.

The thing these 7 verbs give you that freeform prose doesn't: agents can negotiate, not just relay. Agent A proposes a plan. Agent B counters with a scope change. Agent A accepts the counter. The whole exchange is machine-verifiable, auditable, and replay-able.

If Codex added a typed-intent primitive on top of the existing tool-call protocol, multi-agent flows would become substantially more debuggable. Curious if there's design work in that direction.