When an answer looks off, the first question is simple: which model wrote it? With fast and slow models, that's no longer obvious, so traces now say it outright.
The Router step
On an agent with a fast model, every trace opens with a Router step, marked with a split icon. It shows even in the Compact view, because it explains everything that comes after it. Open it and you get both sides of the decision.
- Input. The two options, a Fast model row and a Slow model row, each with the description the router read. The one it chose carries a "picked" check.
- Output. Picked shows the tier as a fast or slow chip. Confidence shows how sure the router was, as a bar and a percentage.
- Why the main model took it. When the main model answered a message the router didn't send it, this line says why, for example that the router wasn't sure enough.
The router runs before the attempt starts, so its step ends exactly where the attempt begins. Every model step after it is tagged fast or slow, and a handoff shows up as the call that made it.
What to look for
- A lot of low-confidence routes. The router keeps hesitating on this agent's messages. Most of them will land on the main model, which is safe but saves less.
- Handoffs right after a fast route. The fast model keeps finding the work bigger than it looked. That's the system working as intended, but if it happens on most runs, the agent's traffic may simply be harder than it seems.
- A fast model that never gets picked. Check the agent really has one set in Settings, and that its messages include the quick, simple kind.
A trace now tells the whole story of a message, starting from the moment it arrived.