Solo, Duet & Double — pick the mode on purpose
When a chat answers you, it can run in one of three shapes. They cost different amounts and they fail differently. Knowing which is which lets you trade cost against stability against quality deliberately — instead of paying for reliability you don't need, or skimping on it where it matters.
The three modes
Solo — one agent answers
A single model handles the turn. It's the cheapest and simplest shape: one call, one bill, one place for something to go wrong. The trade-off is that if that one model has a bad moment — a timeout, a blank answer, a wobble — there's no second chance on that turn.
Use it when load is light and the occasional miss is acceptable: internal experiments, low-stakes questions, anything where a rare retry-by-hand costs you nothing.
Duet — two different agents collaborate
Two different models work the same question and cross-check each other. Because they reason differently, one tends to catch what the other misses — so Duet optimises for answer quality: fewer blind spots, better-rounded answers. It costs more (you're paying for two different brains), and it's a little slower.
Use it when the quality of the answer is what matters most: customer-facing replies, anything where being right and thorough beats being cheap.
Double — two identical instances cover each other
Two identical instances run, and one covers for the other. If the first instance fails or returns empty, the second carries the turn — so Double optimises for stability / failover: the turn survives a single instance going down. It is not trying to improve the answer (both instances are the same model); it's trying to make sure you get an answer.
Use it when uptime matters most: a public chatbot during a busy launch, anything where a dropped turn is worse than a slightly more expensive one.
The crucial distinction
Duet and Double both run "two" — but for opposite reasons, and the difference is the whole point:
- Duet = two different agents → quality. Different reasoning, cross-checked. You pay for diversity.
- Double = two identical instances → stability. Same model, one as a safety net. You pay for resilience.
Cost vs stability vs quality — at a glance
| Mode | What runs | Optimises for | Relative cost | Reach for it when… |
|---|---|---|---|---|
| Solo | One agent | Simplicity / lowest cost | Lowest | Load is light; an occasional miss is fine. |
| Duet | Two different agents | Quality (diverse reasoning, cross-check) | Highest | Answer quality matters most. |
| Double | Two identical instances | Stability / failover | Middle | Uptime matters most; a dropped turn is the worst outcome. |