codex - 💡(How to fix) Fix Shared workspace/message bus for Codex subagents [1 comments, 2 participants]

What variant of Codex are you using?

Codex CLI

What feature would you like to see?

I would like Codex subagents to have an optional shared coordination scope where they can exchange short structured messages, status updates, findings, and handoff notes during a multi-agent workflow.

Current subagent workflows are useful for parallel work, but coordination is mostly parent-mediated: subagents run in separate agent threads and the parent agent waits for results, then consolidates them. This works for independent tasks, but becomes limiting when multiple specialized agents need to coordinate around a shared investigation or implementation plan.

Current behavior

Codex can spawn subagents in parallel and collect their results into one consolidated response. The user can inspect or steer agent threads via /agent.

However, there does not appear to be a first-class shared thread, shared inbox, or common subagent scope where subagents can communicate with each other directly or publish structured intermediate findings visible to other subagents.

Users currently have to emulate shared state manually through files such as:

.agent-work/<task-id>/
  board.md
  decisions.md
  audit.md
  implementation.md
  test.md

That workaround is usable, but it is not first-class, has no built-in locking or ownership semantics, and can cause coordination issues when multiple agents write concurrently.

Proposed feature

Add an optional shared communication scope for subagent workflows.

Possible config shape:

[agents]
max_threads = 5
max_depth = 1
shared_workspace = true
shared_workspace_mode = "append-only" # or "parent-mediated"

[agents.shared_workspace]
path = ".codex/shared-agent-work"
visibility = "workflow" # workflow | project | session

Or a more explicit model:

[subagents.shared_scope]
enabled = true
kind = "message_bus" # message_bus | files | hybrid
visibility = "workflow"
write_policy = "append-only"

Desired behavior

When a multi-agent workflow starts, Codex creates a task-local shared scope, for example:

shared_scope:
  - board
  - agent status events
  - findings
  - handoff notes
  - decisions accepted by parent

Subagents should be able to:

• append status updates
• append findings
• read findings from other subagents
• address messages to another agent or to the parent
• publish final handoff notes

The parent/orchestrator should be able to:

• create the shared scope
• merge findings into accepted decisions
• resolve conflicts
• decide what enters the main conversation context
• summarize shared activity to the user

Example workflow

User prompt:

Use audit, etls, and mcpm in parallel.
They should coordinate through a shared workspace.
audit checks risks, etls checks implementation, mcpm checks MCP/server impact.
Wait for all of them and return a consolidated plan.

Expected behavior:

main thread
  ├─ creates shared scope for this workflow
  ├─ spawns audit
  ├─ spawns etls
  ├─ spawns mcpm
  ├─ agents publish findings/status to shared scope
  ├─ agents can read each other’s published findings
  └─ parent consolidates final result

Why this matters

For real project workflows, subagents are often not completely independent. They may need to coordinate on:

• shared assumptions
• intermediate discoveries
• ownership of files or modules
• whether a finding invalidates another agent’s plan
• conflict resolution before final implementation
• long-running investigation state

Without a first-class shared scope, the parent must manually relay information between agents, or users must create ad-hoc filesystem protocols.

Safety and conflict controls

This feature should not mean unrestricted shared editing.

A safe MVP could be:

• append-only message bus
• parent-owned decisions
• per-agent namespaced notes
• no direct overwrite of another agent’s notes
• optional visibility in TUI
• optional /agent shared or /agent board view
• parent decides what enters main conversation context

For write-heavy workflows, Codex could explicitly recommend separate worktrees or per-agent file ownership to avoid conflicts.

Relationship to existing work

This is related to #12047, which proposes a broader multi-agent TUI overhaul with named agents, shared team inbox, async orchestration, and @mention messaging.

This issue is intended as a narrower MVP-focused request: provide a first-class shared coordination scope/message bus for subagents, even without the full team/TUI model.

Desired outcome

• A multi-subagent workflow can create a task-local shared scope.
• Each subagent can append structured messages/findings/status updates.
• Subagents can read shared-scope entries from other subagents.
• The parent agent can summarize or promote entries into the main response.
• The user can inspect the shared scope from the CLI/TUI.
• The feature is opt-in and backward-compatible.
• Shared writes are safe by default, preferably append-only or parent-mediated.

Additional information

No response