I was planning how Voxelbox, the local agent system I use across my projects, could host interactive Claude sessions. The proposal focused on keeping a better record of a session already underway. I hadn't asked what the host should let that session do.
I handed the proposal to Claude with one instruction: attack it. Find the weakest claim, cite the evidence, and tell me what should stop the work.
Claude found the capability I'd buried. A hosted process could take new input while it was running. Another agent could return a review to the session that requested it. The host could interrupt a process once it had stopped being useful. That gave the system a way to steer the live agent while it worked.
I copied the original proposal and Claude's critique into Codex and asked it to reconcile them. Codex kept one shared session record and the live-control idea, while letting a long-running interactive process take a different path from a batch turn. It also deferred human attach and reconnect. I was going to build a better log; I left with a smaller way to steer a running session.
The system didn't run this exchange. I copied Claude's response into Codex's context and wrote a retrospective afterward (I was the message bus, a role for which I remain wildly overqualified).
Give the review its own job
The original conversation was helping me make the proposal. A critic gets a different assignment: find where the idea breaks and make me answer for it. The same model can do useful review in a separate context because the job has changed. Here I also wanted another perspective. Claude tried to refute the plan; Codex had to reconcile the critique with the implementation constraints. I still had to decide whether the objection changed the plan.
I wanted the decision to survive the conversation: the proposal, the critique, the synthesis, and the reason for the cut. That is the shape I carried into its deliberation design.
I copied messages between models and ended up with less code to write (I've had worse project plans).