Advisor Model

An Advisor Model is a setup where an agent that's working into a hard-to-judge moment calls in a more capable model, gets advice, and then goes back to doing its own work. It's similar to a new employee at a company running into a tricky situation and quickly asking the senior colleague at the next desk for input, then returning to their own tasks. The senior colleague doesn't take over the task — they just offer the judgment needed and step back out.

The advantage of this approach is that it lets you manage cost and performance at the same time. Instead of handing every task from start to finish to an expensive, powerful model, a cheaper model normally handles the work and only calls in the more powerful model briefly at points where the decision is hard. This cuts overall execution costs while preserving the quality of judgment on difficult calls.

The article notes that "a session budget cap and an Advisor Model feature have been added," explaining that this appeared in a previous update alongside a feature that stops execution once the budget runs out. A common misunderstanding is that the Advisor Model doesn't take over and handle the task itself. The original agent remains the one executing the task; the advisor's role is only to add input at the moment judgment gets stuck.