Product Feature Request: Event-Driven Task Coordination for Voice-Led Multi-Agent Work

Thank you for providing voice chat. We really enjoy using it for collaborative software work, and it makes planning, development, testing, and review more natural and productive: the user can inspect the running product, think aloud, refine a design observation in real time, and use the primary chat as a sounding board and coordinator. This is a general work-task coordination pattern, applicable to feature development, UAT, bug fixing, debugging, research, code review, and operational work. It has been especially effective for turning detailed feedback into clear, bounded work without forcing the user to write long issue tickets first.

How we use it today

In our workflow, the primary voice chat acts as the coordinator. The user and coordinator refine a need, observation, or work item; a planning/documentation task can record its scope and acceptance criteria; then an implementation, research, review, or operational task performs the bounded work. The coordinator preserves user control: a task should not expand its own scope or automatically start unrelated work.

Current gap

The current friction begins after a subordinate task starts. Completion is not reliably pushed back to the coordinator, so the user often has to notice that it finished, manually check it, and tell the coordinator to take the next step. Likewise, a worker that needs clarification or is blocked does not reliably interrupt the coordinating workflow with a meaningful event. This creates avoidable pause-and-watch work despite an otherwise highly productive voice-led process.

Instructions to queue future work inside a worker task are also not a reliable substitute for real coordination. A worker should not be expected to interpret and execute a durable sequence of future instructions embedded in its current prompt.

Requested capability

Provide an event-driven coordination model for subordinate tasks. When a task completes, becomes blocked, fails verification, or requires input, it should post a structured event to its originator rather than requiring the coordinator or user to poll for status.

Each event should include:

- Task identity and state: `completed`, `blocked`, `needs review`, `failed verification`, or `waiting for input`.

- A concise summary, relevant artifacts, verification result, and unresolved findings.

- An explicit next-action recommendation, if one exists.

Coordinator-owned queue and decision gates

The primary coordinator should maintain a durable, visible, ordered queue of approved follow-on work. On each returned event, the coordinator evaluates decision gates before dispatching the next item. It must pause the queue when there is a blocker, failed verification, unresolved finding, required user review, or an ambiguity that needs refinement.

Once the gate is cleared, the coordinator can dispatch the next approved bounded task without the user manually watching for completion. This should not mean blind auto-dispatch: the coordinator should surface exceptions and preserve explicit user control where approval is required.

Intended workflow

1. User and coordinator refine a need, observation, or work item until the scope is clear.

2. A planning or documentation task records the scope and acceptance criteria when needed.

3. The coordinator dispatches one bounded implementation, research, review, or operational task.

4. The task posts a completion, blocker, or review-needed event to the coordinator.

5. The coordinator either resolves the exception with the user or advances the next approved queue item.

This capability would make multi-agent coordination more reliable, transparent, and user-controlled while removing unnecessary manual monitoring from an already valuable voice-led workflow.

Thanks,
Niranjan

I’m working with Voice similarly. In my case, the coordinating chat is project-agnostic and delegates work across a large number of local Codex projects. Linear holds the project backlog and current work state.

The failure is usually at the handoff. A project finishes an implementation task or PR, but the coordinating chat does not reliably notice, report the completion, inspect the next available work, or continue through routine steps that do not require my approval. This also happens with text, so Voice does not appear to be the underlying cause.

I use a conservative execution model: subagents are read-only, only one write-capable agent operates at a time, and the coordinator owns the plan, active-work ledger, and final verification.

I would be interested in knowing whether any part of this can currently be made more reliable through global AGENTS.md, config.toml, skills, hooks, or notification settings, or whether reliable completion events and coordinator-owned progression require the product-level capability proposed here.

Josh

Thanks for the detailed workflow. Voice and subagents already support task coordination, but we haven't confirmed the reliable event handoff and durable queue you described: https://learn.chatgpt.com/docs/features/voice. We can pass those specific gaps along, with no timeline to share.