“Why am I copying one agent’s findings into another?”
Let the agents hand off directly.
Coordinator uses native inter-task messages for bounded assignments, results, dependencies, and collision notices when another task actually needs to act.
Stop being the message bus between your AI agents
Make your Codex agents talk to each other. Without letting them step on each other.
Give Coordinator one goal. Separate Codex tasks use native inter-task messages for the handoffs that matter, each gets a clear job, and the work comes back into one result.
- No copy-paste relays
- Clear task boundaries
- Does not copy your chats
01 Give them one goal
The multi-agent problem
Your agents should talk to each other — not through you.
Opening more Codex tasks should create leverage, not turn you into the person copying findings, blockers, and instructions from one window into another.
“It’s hard to keep track of which clone or worktree belongs to which task.”
Use one checkout for everyday work.
Separate tasks can share your existing branch, dependencies, settings, and local tools.
“The main thread and another agent are doing nearly the same work.”
Give every job one owner.
Coordinator reuses related tasks first, then keeps assignments and planned file areas visible.
“I know subagents exist, but I have to remember to ask for them every time.”
Make parallel help part of the instructions.
Subagents are available on supported non-Ultra intelligence levels when requested directly or through project and skill instructions. Ultra adds proactive delegation.
“My code, tests, and docs are drifting across different chats.”
Split one result into complete lanes.
Tasks work from the same current codebase, with each lane and its planned files visible.
“Did one agent just overwrite or undo work from another thread?”
Keep dangerous Git operations out.
Tasks re-read shared files, pause on a real collision, and avoid reset, restore, stash, clean, force-push, and broad staging during parallel work.
Built on native Codex messaging. Coordinator does not invent a second chat network. Codex already provides task threads and inter-task messages; Coordinator gives those messages one shared goal, clear ownership boundaries, and rules for when another agent actually needs to hear something.
What changes
One goal. Agents that can talk. Clear lanes.
Codex does the work and carries the messages. Coordinator keeps the collaboration pointed at one outcome.
-
1
Give them one result
Start with the outcome you want, not a separate prompt for every agent.
-
2
Give each agent a lane
Coordinator reuses useful related tasks first, then gives each durable task one complete job and visible boundary.
-
3
Let them talk when it matters
Assignments, results, dependencies, and real collisions move through bounded native messages; the goal owner brings the work back together.
See the workflow
Ask once. Let the agents work together. Review one result.
Three short views show the workflow. Open optional Mission Control only when you want one read-only overview.
Designed to stay out of your way
Let agents communicate without adding another system to manage.
- No permanent watcher. The goal owner responds when work finishes or needs you. If you request unattended supervision, Coordinator can use one temporary Codex follow-up until the next decision.
- No copied chat history. Codex remains the home for task conversations.
- No message noise. Tasks do not send progress chatter, acknowledgements, or status-check loops; another task gets a message only when it needs to act.
- No always-on dashboard. Mission Control starts only when you ask, reads one project, and refreshes only when you press the button.
- No forced process. Keep the way of working that already suits your project.
Free and open source
Add it to Codex in two commands.
Install the current stable release, open your project, and tell Coordinator the result you want.
1
Install the plugin
codex plugin marketplace add eyeinthesky6/codex-coordinator --ref v0.5.0
codex plugin add codex-coordinator@codex-coordinator
2
Give it a real goal
Use $codex-coordinator for this project goal:
<describe the outcome you want>Installation never turns Coordinator on for every project. You choose where to use it.
Built in the open
Try it on a real multi-agent workload.
If you have been copying findings between Codex tasks, checking every window, or untangling duplicate work, give those agents one goal and let Coordinator handle the handoffs.