How to Create a Multi-AI Team in aiDex

An AI team in aiDex is a saved panel of up to five named personas, each pinned to its own model, that answer together while a moderator watches for agreement. This guide shows you how to create one step by step, choose a model for each persona, and run it. For the bigger picture, see Multi-Model AI Workflows.

What is an AI team, and when should you use one?

An AI team is a reusable group of distinct AI personas that respond to the same conversation at once, with a moderator model that tracks where they agree. Each member is a model plus a written persona prompt, so one team might pair an optimist, a skeptic, and a domain expert, each on the model that suits its role.

Use a team when a question benefits from more than one viewpoint: pressure-testing a plan, reviewing a decision from several angles, or getting domain-specific reads (legal, financial, technical) in one thread. For a quick fact or a routine draft, a single model in Solo mode is faster and cheaper. A team earns its cost when the disagreement and the synthesis are the point.

Step 1: Go to the teams page and name your team

Start at Teams to see your saved teams, then open the team builder to build a new one. The first thing you set is the team name. Pick something that describes the team's job ("Product Review Panel", "Contract Skeptics") rather than the models inside it, because the models can change but the role usually does not. The name is how the team appears when you open aiDex and in the sidebar, and opening a saved team starts a Team-mode chat directly.

Step 2: Add up to five personas, each on a chosen model

Add your members one at a time, up to five. Each member has two parts: the model it runs on, and a persona prompt you write. The model is the engine; the persona prompt is the job description. Give every persona a distinct role or stance so the panel covers different ground. A few patterns that work well:

• A proponent and a critic, so one argues for an idea and one argues against it.
• Specialists by domain, such as an engineer, a designer, and a finance lead.
• A range of temperaments, like a cautious risk reviewer next to an ambitious strategist.

Avoid stacking five personas that would all answer the same way. If two members would say the same thing, you are paying twice for one opinion. The value of a team comes from deliberate contrast.

Step 3: Write persona prompts that give each member a real role

A persona prompt is the short instruction that tells one member who it is and how to behave. Write it in the second person and be specific about the role, the stance, and what to focus on. "You are a security reviewer. Look for ways this plan could fail or be abused, and name the single biggest risk first" gives the model a clear job. "You are helpful" does not.

Good persona prompts do three things: assign an identity, set a point of view, and define what the member should pay attention to. The sharper the contrast between your prompts, the more useful the panel's spread of answers will be. You can revise these later, so start with a clear first draft and tune it after you see the team respond.

How do you pick the right model for each persona?

Match the model to what the persona is for. aiDex supports OpenAI, Anthropic (Claude), Google (Gemini), DeepSeek, and local models through Ollama, and they have different strengths, so a reasoning-heavy critic and a fast brainstorming voice do not need the same model. For a full breakdown, see Which AI Model for Which Task?.

A practical approach: put your most demanding role on a stronger model, and let lighter or higher-volume roles run on faster, cheaper ones. Mixing providers across personas is a feature, not a problem: it is the whole reason a team beats five copies of one model.

Step 4: Pick a moderator model and save

Choose a moderator model, then save the team. The moderator is a separate model that reads the members' responses and watches for consensus rather than adding its own persona answer. The default is a fast Claude Haiku, a sensible choice because the moderator runs often and does not need to be your most expensive model. You can pick a different one for stronger synthesis, but for most teams a fast, inexpensive model is the right call. Once you save, the team is ready to open from aiDex or the sidebar.

How does the moderator and consensus work?

The moderator is the team's coordinator: as members respond, it watches the conversation and looks for where they agree, so you get a sense of consensus instead of disconnected answers to sift through yourself. That turns a pile of opinions into something you can act on, and keeping it on a fast model (the default Haiku) keeps the team responsive while your members do the heavier thinking. To dig deeper, see How to Get a Consensus Answer.

Step 5: Run your team in Team mode

Open the saved team from aiDex or the sidebar to start a Team-mode chat. Type your prompt as usual, and the members each respond in their persona while the moderator tracks consensus. You are talking to the whole panel at once, not one model at a time.

Treat the first run as a test. If a persona drifts off-role or two members overlap too much, go back to Teams, adjust the prompts or swap a model, and save again. Teams are meant to be tuned over time as you learn how your roles behave.

BYOK or managed credits: which should you use?

Use your own provider keys or the ones we manage, and pick the models you want. With BYOK you bring your own keys for OpenAI, Anthropic, Google, DeepSeek, or local Ollama models and pay your providers directly. Managed credits let you run on aiDex's server-side keys without setting up your own, the simpler path if you do not already have provider accounts. Since a team can call several models per turn, it is worth knowing your setup, but either way the team-building steps above are identical.

Next steps

Once your team runs cleanly, push it further. If you want your personas to argue a position rather than just answer, see How to Run a Debate Between AI Models. And when you need the panel to land on one clear recommendation, see How to Get a Consensus Answer.