Here's the workflow I kept doing manually: open Pane, create three panes for three issues, start Codex in each one, wait for them to finish, open review tabs, send each reviewer /review, copy findings into the implementation tabs, tell the implementation agents to fix the real issues, check results, prepare PRs.
Every step was a click or a command I typed. And every step was something an agent could do for me. So we're building Pane Chat: one conversation that handles the whole thing.
what pane chat is
Pane Chat is a global orchestrator terminal inside Pane. Normal agent tabs sit inside a pane and work on one repo, one branch. Pane Chat sits above all of that. It can see every repo, every pane, every agent. You talk to it, and it coordinates the rest.
It uses the same runpane CLI that any external agent can use. There's nothing proprietary about it. Pane Chat is just an agent with a terminal, running runpane commands, same as an orchestrator script would.
orchestrator, not implementer
The key decision was: Pane Chat coordinates, it doesn't implement. When you ask it to build a feature, it doesn't start writing code itself. It creates a pane, starts an implementation agent in that pane, and checks on its progress.
Why? Context. An agent that plans, implements, reviews, and tests all in one conversation will fill its context window and lose track of earlier decisions. Splitting the work across agents gives each one a clean context focused on one job.
This is the same pattern Anthropic describes in Building Effective Agents: keep orchestrators lean, delegate to specialized workers, let each one start fresh.
the workflow
Here's what a real session looks like:
1. I tell Pane Chat: "There are three issues that need work. Create a pane for each, start Codex, and tell me when they're ready for review."
2. Pane Chat discusses the approach: "Issue 42 touches the auth module, I'll scope the prompt to that. Issue 43 is a new endpoint, I'll give it more room. Issue 44 is a bug fix, quick scope."
3. It creates three panes with runpane panes create, each with a tailored initial prompt.
4. It waits for each agent to finish, reads their output with runpane panels screen, and reports back: "42 is done and clean. 43 needs your input on the API schema. 44 is still running."
5. After implementation, it opens review tabs in each pane, sends /review, waits for findings, and tells me what's real vs noise.
local skills, not giant prompts
Pane Chat's startup prompt is one line:
Use the pane-orchestrator skill and initialize yourself as Pane Chat.The actual workflow lives in local skill files that Pane caches in the background. This keeps the initial context small and means the orchestration logic can evolve independently of the agent's training data.
claude or codex
Pane Chat works with Claude Code or Codex. A toggle in the top-right switches which one runs the orchestrator. Both use the same runpane commands. The only difference is which agent you're talking to.
This matters because orchestration is a skill some agents are better at than others. Try both. Switch when one handles your workflow better.
where we are
Pane Chat is in active development. The core loop works: discuss, create panes, delegate, check results, report. We're still refining the skill files and the delegation patterns. If you want to try it, update Pane and look for Pane Chat in the interface.
The orchestrator skills are open source: Codex and Claude Code. Fork them, change the workflow, make them yours.