Adaptive routing
Predicting per-call difficulty, and paying for the expensive model only when it is needed.
Most calls in an agent loop are not hard. Reading a file, running a test, applying a small edit — these do not need a frontier model. Adaptive routing predicts the difficulty of each call and sends easy work to a cheaper tier.
Two signals
- Intrinsic — extracted from the issue text itself, with the prompt scaffolding stripped out
- Struggle — read from the conversation tail: failing test output, turns burned without progress
The scaffolding-stripping matters more than it sounds. The first version of the predictor scored the whole message, but harness prompts are large by construction, so every call saturated at "hard" and the mechanism degenerated into always-expensive — costing about twice the baseline. That failure is what produced the current design.
Escalation does not stick
After an expensive call, routing drops back to the cheap tier unless the struggle signal persists. A single hard step does not condemn the rest of the task to the expensive tier.