Every thread runs in one of three modes. They are not settings on a slider — they produce genuinely different conversations, cost different amounts, and are each wrong for something.
The short version, if you only read one paragraph: Conductor for most things, Sequential when you want the panel to argue, Parallel when independent agreement is the evidence you need.
The panel
Four models, each with a job the prompt actually assigns it:
| Model | Asked to |
|---|---|
| GPT | Structure the problem and the decision criteria |
| Gemini | Synthesise long context and reconcile competing views |
| Kimi | Develop practical alternatives and verify implementation details |
| Claude | Challenge assumptions and inspect the quality of the reasoning |
Those roles are not labels on a diagram — they are in the system prompt each model receives, which is why the answers differ in kind and not just in wording. You can switch any of them off or replace a seat; Sequential follows the panel's seat order.
Conductor — a model decides how much panel the question deserves
The default. Before anything else happens, one model reads your request and makes a judgement: can I answer this myself, or does it need several viewpoints?
Most requests get answered on the spot. "What is the English name for 水豚?" costs one model call, and calling three more to also say "capybara" would cost you four credits for one fact. When that happens the answer is tagged Direct answer and carries a note saying the rest of the panel was not called — because "four models selected, one replied" looks like something broke, and it isn't.
When the question does warrant a panel, the conductor writes a plan: one to five steps, each assigned to the model best suited to it. The plan is a real message in the thread, not hidden state. It is stored, it is billed like any other call, every later step reads it, and it survives a reload. The turn that wrote it carries a Planner badge.
Then each step runs in order, and each assigned model can see everything above it.
What it is good at: not spending your money. On a mixed workload it is roughly a third the cost of running everything through a full panel, because most messages genuinely are one-model questions.
Its failure mode: the conductor under-calls the panel. It is biased toward restraint, and sometimes you disagree.
The override: ask. "I want each model to answer", "compare your views", or naming specific models all outrank the conductor's judgement — the instruction to honour an explicit request sits at the very top of its prompt, above all the cost restraint. If you ask for the panel and get one answer, that is a bug, not a policy.
One thing the conductor is explicitly forbidden from doing: writing anything attributed to another model. If it has not called Gemini, it may not write a paragraph headed "Gemini:" summarising what Gemini would have said. A fabricated panel is worse than a missing one.
Sequential — each model reads the ones above it
Every selected model answers in seat order, and each one is given every answer before it as context.
This is the mode that produces an argument. Model two is not answering your question in isolation; it is answering it having just read model one, with an instruction to add new signal, correct weak claims, and make the decision more defensible rather than repeating what is already there.
For project work, Sequential is a review chain with one committer. Every model before the last may list, search, and read the real project and may use web research, but it cannot write, create, or delete files. Only the final model receives modification tools; it reads the complete discussion and applies one coherent implementation to the shared project. This avoids a trail of intermediate edits being repeatedly overwritten while preserving evidence-based reviews.
Order matters here more than anywhere else. The last model has the most context and the best position to synthesise or to demolish. The seat design puts the structural critic last; replacing a seat changes its occupant without creating a second, contradictory order setting.
What it is good at: decisions where you want the disagreement surfaced rather than averaged. Four models that have read each other and still disagree have told you something a single confident answer never could.
Its failure mode: later models can anchor on earlier ones instead of thinking independently. If you suspect that, run the same question in Parallel and compare.
Cost: one credit per model, every message. Four models is four credits.
Parallel — nobody sees anybody
Every selected model answers the same question at the same time, with no knowledge of the others. Their answers stream side by side in stable seat order.
What it is good at: independence. Agreement between models that never spoke is real evidence; agreement in Sequential mode might just be politeness. Use it when the question has a checkable answer and you want to know whether the field converges.
Its failure mode: repetition. Four models independently answering an easy question produce four similar paragraphs and cost four credits.
Code remains independent too: every model receives file tools backed by its own private copy of the same fixed project baseline. Generation never writes the shared workbench, so there is no last-writer-wins race. Each completed result carries a baseline diff; choose one model's card to replace the project with that complete candidate. Switching models always rebuilds from the baseline-derived snapshot and never performs a three-way merge. If the project changed in the meantime, the app asks before discarding those changes.
Choosing, in practice
| You want | Mode |
|---|---|
| An answer, cheaply | Conductor |
| A plan carried out across models | Conductor |
| The panel to argue and build on each other | Sequential |
| To know whether models independently agree | Parallel |
| One reviewed implementation written to the project | Sequential |
| Independent code candidates to compare | Parallel |
| A second opinion on a Sequential result | Re-run in Parallel |
The settings that actually change outcomes
Which models are on. Click a model anywhere it appears — the row under the thread title, or the rail beside the composer — to switch it in or out. Both places are the same switch and both persist immediately.
Their slots. The settings button in the composer opens one flat model list. It deliberately does not expose Advanced, Complex, or Simple classifications. Replacing an occupant still preserves its internal task and Sequential position.
All four models start enabled in every mode. You can switch any one off for a thread, and that explicit choice persists when modes change. Gemini and Kimi are therefore available to the Conductor from the first request instead of being hidden in the optional model pool.
What a run costs, and how to stop one
One credit per completed model answer. A Conductor direct answer is one credit; a four-model Sequential message is four; a Conductor plan with two steps is three (the plan is a call too). An answer that fails is not billed.
While a run is going, the send button becomes Stop. Stopping is a real request to the server, not a closed browser tab: the signal reaches the provider, tokens stop being generated and stop being billed, and no further model is called. Whatever text had already arrived is kept in the thread with a [Stopped] marker — you paid nothing for it, and you do not lose what you had already read.
A suggested first hour
- Ask something real in Conductor and watch it answer directly. That is the mode working, not failing.
- Ask something genuinely contested — a trade-off with no clean answer — and watch it write a plan.
- Switch to Sequential, choose the models you want active, and ask the same question. Read the last answer first.
- Run it again in Parallel. If all four still land in the same place, you have your answer. If they scatter, the question was underspecified — which is also worth knowing.
- Press Scribe in the side panel. One model reads the whole thread and writes where it currently stands: the verdict, its confidence, the live disagreement, and what to do next.