How it works — the four steps

What happens between your prompt and the combined answer.

1

Step 1: Figuring out the roles

Before the AI models are asked anything, the system works out which expert roles the task needs and frames a focused instruction for each one.

Each expert prompt is also conditioned on your jurisdiction, so every model applies the right standards and legislation.

2

Step 2: Asking multiple AI models in parallel

Every request is sent to multiple AI models in parallel, one copy per role, so you get independent answers instead of a single model's view.

Analysis
For questions and analysis. Models answer once. Available now.
Build & Verify
For code and other work that can be tested. Generates tests, builds, checks, repeats — with safety limits. Available now.
Creative
For brainstorming and naming. Produces diverse ideas, rates them, you choose. Coming soon.
3

Step 3: Comparing and combining the answers

After the models respond, the system measures cross-vendor consensus and produces a synthesized answer where they agree — flagging where they diverge.

4

Step 4: Multi-stage projects (optional)

For larger projects with several stages, the system runs the above repeatedly, one stage at a time, with a review checkpoint between stages. This comes in a later release.

🔌

How the orchestrator talks to each model

The orchestrator never calls a model's API directly. It speaks one generic internal interface, and a small Bridge per model translates that into each provider's specific API.

Orchestrator
speaks one generic interface
Bridges
one adapter per model
Claude API
GPT-5 API
Gemini API
Perplexity API

Why this matters:

  • Adding a new model means writing one new Bridge — nothing in the orchestrator changes.
  • The privacy filter, token accounting, and token conversion all live in the orchestrator, above the Bridges, so they apply uniformly to every model.
  • Provider-specific quirks (auth, streaming, rate limits, token reporting) are contained inside each Bridge.