Published: 2026-06-26
Analysis & perspective

Open Engine: A Shared Queue to Make Claude, ChatGPT & Codex Hand Off Work

Chapters / key moments (click to jump — plays here on the page)

Most AI-productive people juggle five or six tools — Claude for front-end, OpenAI/Codex for back-end, OpenClaw or Hermes for autonomy — and become "the hallway" that carries work between them by hand. Nate B Jones's answer is Open Engine: put the work in a queue that both people and agents can read and write, so a ticket — not a chat box — becomes the shared place where any agent or teammate can claim, work, and hand off a task with full state attached.

Source video

"I Built an Open Engine That Connects Claude, ChatGPT, and Codex Together" by Nate B Jones — Watch on YouTube →

Key Takeaways

The bottleneck is the boundary, not the model. Claude and Codex don't do the same things, so you trade between them — the pain is the handoff: getting work to leave one tool with its sources and land with the right next agent or person, without anyone reading a giant transcript.
A ticket beats a prompt. A prompt asks for an answer; a ticket asks for a result and carries owner, background, what the agent may do, where it must stop, and what it has to show when done — so multiple agents that don't know each other can collaborate through the ticket as their shared workspace.
Five components. Open Engine is a Linear queue (chosen for its generous free plan; Jira or your own board also works) plus four skills that teach the AI the protocol: a setup skill, a status skill, a run-the-queue skill, and a smoke test. You can point OpenClaw or Hermes at the same skills.
The claim-lock loop. An agent checks its assigned queue, claim-locks an eligible "agent-instructions" issue, moves it to agent working, leaves an "agent claimed" receipt, does the work locally, then moves the issue to done — so you can audit what happened instead of asking "did you do it?"
Cross-provider, cross-team delegation. One person's Codex can write a self-contained issue and assign it to another person's Claude agent, which picks it up on its own heartbeat. If an agent hits ambiguity it moves the ticket to needs input and asks the exact blocking question rather than guessing — keeping the audit trail in one place.
Smoke test first. The clean validation is to create an issue called "say hello," assign it to a human or agent with the instruction label, and confirm it can move to done — proving the interaction loop works before you load it with real work.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Use a queue both humans and agents can use
The pattern is a ticket queue an agent can write to and you can read: the demo uses Linear (free plan) but says Jira or any board works. Each issue states the outcome, the owner, the background, what the agent may do, where it should stop and what it must show when done.
Teach each agent the protocol with four skills
His guide has five parts in all: the queue plus a setup skill, a status skill, a skill that runs the queue, and a smoke test. You can also point OpenClaw or Hermes at it.
The claim-and-receipt loop
An agent wakes on its heartbeat and checks its assigned queue for issues labelled "agent instructions", claim-locks one, moves it from agent to-do to agent working, and leaves an "agent claimed" receipt before starting. If it hits ambiguity it moves the issue to needs input and asks the exact blocking question instead of guessing; when finished it leaves agent done plus proof.
Cross-agent handoff
One person's agent can write a self-contained issue for another person's agent on a different model provider, assign it, and label it; the audit trail stays on the issue instead of in chat.
Smoke test
Create an issue called "say hello from the queue", assign it to a person or agent with the instruction label, and confirm it can move to done.

Gotchas

The guide and skills are shared through his Substack community, so the exact files aren't shown on screen.
He says a chat box is a poor place to keep state; the point is a system of record with receipts rather than a smarter agent.