How do I work with two AIs together?
Facts as of October 17, 2026
In detail
The basic rule: a main AI decides and does everything it can. What it does less well, such as images, it hands off to a helper AI. The main AI sets up the helper AI itself and operates it. That way you don't have to maintain two AIs and two sets of settings, but one. Which AI is suited as the main AI is covered under Which AI is the right one – and what do the others offer?
For the handovers between the two:
- Fixed roles. The main AI plans, checks and keeps records; the helper AI produces what the main AI does less well. You decide and approve. The cycle: task, approval, dated handover, check, updated status in the knowledge store.
- One place keeps the status. What is open is in one place. If the executing AI finds a contradiction with it, it reports it instead of resolving it itself.
- Handovers dated and separate. Every handover is a file of its own with a date; earlier ones are not overwritten. Otherwise two contradicting versions sit side by side, and which one applies can only be guessed from storage location and timestamp.
- Check abilities, don't assume them. Announcing access and actually having access are two different things, and no AI says so on its own in advance. You notice when the first file won't open; a single question beforehand would often have been enough.
- Your eyes count. A finding from machine reading of an image is a reason to ask, never a reason to act.
- Only what changed comes back. After a cross-check, the list of errors goes back as a file, with the request to deliver only the changed files again.
In practice, this means for the helper AI:
- It reads only what the main AI lays out for it. That is a copy of the current status, which the main AI maintains and checks against the original. If something is missing, the helper AI asks instead of looking for it elsewhere, filling it in from memory or making it up.
- It writes to only one place, an inbox of its own. From there, the main AI takes over what applies. Whatever the helper AI changes gets a new name; it doesn't touch the original.
- Every round ends with a new handover, dated, never overwritten: what was done, what you approved, which files belong to it.
- No list of open questions of its own. Only the main AI keeps that. The helper AI reports contradictions in the handover instead of resolving them itself.
- Its knowledge stays at the status of the last handover. It doesn't know what has changed since then until the main AI tells it. Every new handover therefore begins with what has changed and says explicitly that it replaces the old status.
- Test permissions, don't assume them. Whether the helper AI can read or write in a given place is shown by a trial run with a test file before the first real task comes.
- What the main AI can do itself, it takes back. The helper AI does only what it demonstrably does better. Analyzing, keeping lists, writing messages and checking stay with the main AI, even if they were once outsourced.
In practice
For a talk in October 2026, Thomas J. Weiss had the slides built like this: Claude wrote the brief as a long text file – slide texts, speaker notes, image ideas, guidelines for style and character. ChatGPT laid it out and drew. Claude checked the result in a separate role as a cross-checker from the audience's perspective and found eight documented errors. The second cross-check found all points implemented and one more error, which had already been in the previous version and had been overlooked.
While preparing the performance of a play for his parish in September 2026, at times three AIs ran side by side. The chronicle recorded that the handover between two AIs took more care than the work being handed over. One day, one AI built a large document to hand over images to the other; but that route only transferred text. A question beforehand would have been enough. The chronicle notes: a person who has read both chats sees the difference immediately.
In October 2026 it became clear why the helper AI should only read what the main AI lays out for it. It had been reading in a backup copy that didn't keep up with replaced files, and found an old version there; its own status hadn't been renewed since the beginning of September. Since then, every handover to it begins with the sentence saying what has changed and that it replaces the old status. Its permissions were tested with a trial before its first use, and the main AI took back tasks it can do itself.
On October 17, 2026, he turned this into a fixed rule: Claude is his main AI, and he now uses ChatGPT only for images. Claude sets up ChatGPT itself and operates it.
But it's clear, there's one AI that decides, and one AI to which some tasks are outsourced from the other AI. That's how I've built it by now.
(Original: „Aber es ist klar, es gibt eine KI, die entscheidet, und eine KI, in die manche Aufgaben von der anderen KI ausgelagert werden. So habe ich es inzwischen gebaut.“)
Even before that, he wanted as little back and forth as possible:
In order to limit the back and forth and the setting of conventions and the technical familiarization and so on, I would want to work with only one AI. I might switch AIs if I hoped for a much better result from another one, and that also demonstrably proved true.
(Original: „Ich würde, um das Hin und Her und das Festlegen von Konventionen und das technische Einarbeiten und so weiter zu beschränken, nur mit einer KI arbeiten wollen. Ich würde unter Umständen die KI wechseln, wenn ich mir von einer anderen ein viel besseres Ergebnis erhoffe, und sich das auch nachweislich bewahrheitet.“)
Sources
- Note „Wie dieser Vortrag entstand“, replicapong.de, 11.10.2026
- Chronicle of a theater project, 04.09.2026, with addendum from the same evening
- Way of working of the same project, as of 26.09.2026
- Interview on the review of the same project, 02.09.2026
- Dictation by Thomas J. Weiss comparing with the theater project, 17.10.2026
- Handover texts to the ongoing chats, 07.09.2026
- Handovers to the helper AI in the same project, 16.10.2026
- Chronicle of the same project, 16.10.2026
- Template for the helper AI's access, 16.10.2026
- Start prompts for the helper AI in the same project, 17.10.2026
Cite: Thomas J. Weiss: “How do I work with two AIs together?” replicapong.de, as of October 17, 2026. https://replicapong.de/en/question/two-ais/ – CC BY 4.0