Choosing the model, or several at once

Every agent thinks with an AI model — a particular Claude, GPT, Gemini, Grok or local model, at a particular depth. Think of models as the agent’s brains: a bigger one reasons harder and catches more but takes longer; a lighter one answers fast. The agent stays the same whichever brain it uses (see Talking to agents).

Each agent already has a sensible default, so you never have to think about this.

Switching the model

The small gear button just to the left of the message box opens the model menu. It lists the models your box offers, with the current one checked. Pick a different one and your next message goes to it. Your pick sticks until you change it, and it belongs to this conversation: your other conversations, and the agent’s default everywhere else, stay as they were.

Less common models are folded under More so the list stays short.

Asking several models at once

For a question that matters — a big decision, a contract, a tricky diagnosis, or anything where you’d want a second opinion — check two or more models in the menu before sending.

The agent then thinks about your message on each model independently: each one sees your words and the conversation so far, but not the others’ answers, so when they agree it actually means something. When they’ve all finished, you get one reply in the agent’s voice that says where they agree, what only one of them caught, and whether they split.

Expand the reply to see each model’s own answer, and expand again for its raw work.

Debate

If the models disagree, the reply’s footer says they split · Debate. Press Debate inside the expanded reply (optionally with a line saying what to focus on) and each model is shown the others’ answers and asked to respond. The reply is then rewritten as the whole story — what was considered, what was argued, and what it concluded — with earlier versions kept in the expansion. You can Debate again a few times. Debates only happen when you ask; you can also just tell the agent “have them argue it out.”

Keeping one answer

If one model’s answer is the one you want, press Use this answer on it inside the expanded reply. That answer becomes the reply, the other models’ answers are dropped from the record (and from those models’ memory, so they don’t argue for them later), and if a slower model is still working it is stopped. You can Undo until the next message. You can also just tell the agent “go with Claude’s.”

If you simply send your next message while a model is still working, it is stopped and the reply is made from the answers that finished. The models all work in the same copy of your project, so anything one of them edited is already there.

When to use it

Use several models for judgment on what’s in the conversation: decisions, reviews, second opinions. For getting work done, a single agent handing jobs to workers is usually better (see Talking to agents). Several models cost more, so save them for the questions where being wrong would cost more.

Picking a good one

You don’t need to know the model names. The rule of thumb:

Hard, careful, or high-stakes work — reach for a heavier model. “Use the strongest model for this one” is a perfectly good instinct.
Quick, simple, or throwaway work — a lighter model gets you there sooner.

If you’re not sure, leave it on the default. It’s chosen to be a good all-rounder.

Fast mode

Some models offer a fast mode — a toggle at the bottom of the model menu, shown only when the selected model supports it. Turn it on and that engine prioritizes getting words on the screen quickly.

Staying in control

Changing the model never changes who the agent is or what it’s allowed to do — only how it thinks. Your admin (or the Lowkey team) decides which models are offered; ask them if one you want is missing.